Renewable and Sustainable Energy Reviews 39 (2014) 587–603

Contents lists available at ScienceDirect

Renewable and Sustainable Energy Reviews

journal homepage: www.elsevier.com/locate/rser

Optimisation of energy management in commercial buildings with

weather forecasting inputs: A review
Dimitris Lazos n, Alistair B. Sproul, Merlinde Kay
The University of New South Wales, Australia

art ic l e i nf o a b s t r a c t

Article history: Information about the patterns that govern the energy demand and onsite generation can generate
Received 5 March 2014 significant savings in the range of 15–30% in most cases and thus is essential for the management of
Received in revised form commercial building energy systems. Predominantly, heating and cooling in a building as well as the
26 May 2014
availability of solar and wind energy are directly affected by variables such as temperature, humidity and
Accepted 6 July 2014
solar radiation. This makes energy management decision making and planning sensitive to the prevalent and
future weather conditions. Research attempts are being made using a variety of statistical or physical
Keywords: algorithms to predict the evolution of the building load or generation in order to optimise the building energy
Weather forecasting management. The response of the building energy system to changes in weather conditions is inherently
Load forecasting
challenging to predict; nevertheless numerous methods in the literature describe and utilise weather
Distributed generation
predictions. Such methods are being reviewed in this study and their strengths, weaknesses and applications
Energy management
Commercial buildings in commercial buildings at different prediction horizons are discussed. Furthermore, the importance of
considering weather forecasting inputs in energy management systems is established by highlighting the
dependencies of various building components on weather conditions. The issues of the difficulty in
implementation of integrated weather forecasts at commercial building level and the potential added value
through energy management optimisation are also addressed. Finally, a novel framework is proposed that
utilises a range of weather variable predictions in order to optimise certain commercial building systems.
& 2014 Elsevier Ltd. All rights reserved.

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 587
2. Forecasting techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 589
3. Weather variable prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 590
4. Generation forecasting methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 592
5. Load forecasting. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593
6. Energy management optimisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 596
7. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 598
8. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 600
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 600

1. Introduction growing number of commercial buildings. Such measures are

aiming to either generate energy onsite or manage the demand
As a response to increasing electricity prices and greenhouse via energy efficient designs, upgrades and demand response (DR)
gas emissions, a variety of measures are being adopted by a policies; often a combination is present. These energy systems are
subject to high degrees of optimisation in terms of financial
savings as the value of energy is variable and affected by a range
n
Corresponding author.
of factors, such as the time of day, season and the energy source.
E-mail addresses: [email protected] (D. Lazos), Hence, there is a substantial research aiming to achieve optimal
[email protected] (A.B. Sproul), [email protected] (M. Kay). energy performance of commercial buildings. Ideally, during

http://dx.doi.org/10.1016/j.rser.2014.07.053
1364-0321/& 2014 Elsevier Ltd. All rights reserved.
588 D. Lazos et al. / Renewable and Sustainable Energy Reviews 39 (2014) 587–603

Fig. 1. Average weekly load over 2009 for a university building with a large HVAC system in Spain [87].

Fig. 2. Commercial load profile features on a typical weekday [2].

periods of high energy cost, consumption should be minimised

and generation, where available, should be maximised. Thermal or
electrical storage are frequently used as energy buffers and
regulators of demand during the day. The key to such optimisation Fig. 3. Correlation of daily peak load and average temperature for electrical
algorithms is analysing the system's behaviour and predicting its conditioned and gas heated commercial buildings in Sydney [5].
evolution as accurately as possible on a rolling horizon of some
minutes up to several days ahead, so that generation and demand affected by the prevalent weather conditions. The dependence of
can be matched and produce the maximum potential savings. peak load on temperature for a range of commercial buildings in
Commercial building loads are distinctive as they generally Sydney is illustrated clearly in Fig. 3. Additionally, distributed
vary in regular diurnal, weekly and seasonal patterns. Moreover, generation (DG) from renewable energy sources (RES) are inher-
they are to a large extent related to the weather and provide ently tied to the weather. Therefore information about the evolu-
grounds for optimisation as long as the governing patterns are tion of weather variables are valuable input data for optimisation
well understood [1]. On a daily basis, notable similarities appear algorithms.
between typical commercial building energy profiles: the presence Both residential and commercial loads are equally significant
of a base load, a morning ramp-up, an afternoon peak followed by contributors to the energy consumption mix globally. Compared to
a “shoulder” and finally an evening recession towards the base residential building loads, commercial load energy management
load [2]. Weekend load profiles are usually reduced in magnitude. has certain features that are drivers for research in the field of
A typical commercial building weekly load can be seen in Fig. 1. forecasting and optimisation using weather variable inputs. The
Mathieu et al. proposed a characterisation of the features of a main reasons for focusing on commercial buildings in this review
regular weekday commercial energy consumption profile, which is are summarised in Fig. 4. These reasons have resulted in the
displayed in Fig. 2. Depending on the local climate, prominent development of a multitude of approaches to load forecasting and
variations may be observed between seasons. Typically, major energy management [3]. Zhao and Magoules [6] reviewed the
efforts have been made in predicting and optimising the thermal recent trends in forecasting building energy consumption with
aspects of building energy systems, because heating, ventilation particular emphasis on machine learning techniques. Foucquier
and air conditioning (HVAC) accounts for the largest part of the et al. [7] complemented their effort by expanding the list of
overall energy demand in commercial buildings [3,4]. Evidence algorithms that appear in the literature, comparing and classifying
also shows that HVAC is majorly responsible for the magnitude of them. Another recent effort by Sun et al. [8] focused on examining
peak loads [5]. Total and peak HVAC loads are predominantly methods of shifting peak load in buildings.
 D. Lazos et al. / Renewable and Sustainable Energy Reviews 39 (2014) 587–603 589

autoregressive (AR) and a moving average (MA) model. The AR and

MA parts are responsible for linking the present value of the time
series to its past values as well as some past random error
respectively. When there is an integrated part added to ARMA
methods, which allows removing any non-stationarity from the
data [12]; in such cases the models are referred to as ARIMA. The
AR family is capable of simulating the processes that govern the
evolution of variables that are subject to randomness (stochastic
forecasts).
Stochastic forecasts are often combined with deterministic
models that can predict the evolution of variables that are
independent of random error. Exponential weighting (EW) is an
example of such a technique. The principle of EW approaches is
assigning appropriate weights to observations in a time series in
order to develop forecasts for future points. The Holt–Winters
(HW) approach [13,14] is of particular interest to energy con-
sumption forecasting as it accounts for seasonality patterns in a
time series that are linked in an additive or more commonly in a
multiplicative manner. The presence of seasonality in variables
such as ambient temperature or load allows the use of Fourier
time series as an alternative method to approximate the wavelike
behaviour of the variable.
A different approach in forecasting is realised in the form of
Fig. 4. Comparison of residential and commercial building energy management
algorithms simulating learning processes, mainly in the form of
features that are relevant to weather variable based optimisation.
neural networks (NN). Rather than using the input data to
decompose the time series and develop a fit using a number of
With the realisation of the importance of weather variable parameters, NN attempt to simulate the non-linear and non-
inputs in the energy management of commercial buildings, this stationary univariate or multivariate dependences through net-
paper attempts to review methods for prediction of weather works similar to those found in the central nervous systems of
variables and discuss their contribution to the prediction of mammals. The network consists of layers of nodes that form a
generation and load, as well as the potential to add value to number of staged connections of different weights between inputs
building energy management. The first section deals with studies and outputs. Training data are fed in the input layer and evolve as
dedicated to locally predicting temperature or other relevant they propagate towards the output layer to match the actual data
weather variables for purposes of building energy management. as closely as possible. This is mainly achieved by an algorithm
Generation and load prediction methods that use inputs from local known as back propagation; the parameters of each node are
or external weather forecasting are reviewed and assessed in the being continuously modified according to a stream of feedback
following sections. Finally, the outputs of the aforementioned that flows backwards and describes how accurate the configura-
methods are serving as inputs for optimisation algorithms, which tion is compared to the target results [15]. Support vector
are discussed in the last section. The rationale behind this machines (SVM) are promising machine learning alternatives
sequential layered structure serves the aim of linking the fore- utilised in forecasting in the context of commercial buildings.
casting techniques proposed in the literature to a holistic energy SVM attempt to model non-linear relationships based on a
optimisation framework based on weather variable inputs. structure risk minimisation principle that aims to minimise the
upper bounds of error of the object function. Their distinct
advantage over NN is the ability to locate global minima rather
2. Forecasting techniques than local minima in the solution space as well as the ability to
solve non-linear systems with a smaller training dataset [16].
Before proceeding to the discussion of the applications of Physical forecasting methods are based on attempting to
weather, generation and load forecasting in the energy manage- mathematically model the physical processes that characterise a
ment of commercial buildings, it is deemed necessary to present system in order to predict its future state. In the case of buildings,
and compare the features of the most common groups of fore- the physical processes typically include both its structural and
casting techniques used in the literature. Broadly speaking, fore- thermodynamical aspects, as well as interactions with its internal
casting can be classified into statistical, machine learning and and external environment. While historical data are typically not
physical methods. Hybrids are not uncommon in the literature, as required, the amount of inputs differs from case to case. Hence
each group contains features that can complement each other and there are varying degrees of parameterisation and complexity of
improve forecasting quality. physical models. According to the amount of details known and
Statistical forecasting approaches are based on analysing a set accounted for, physical models are known as white box for high
of data points in a time series of the variable in question. The time number of inputs or grey box for a lesser amount of inputs. For
series regression can be used to forecast the future values of a buildings, these techniques often allow the development of
particular variable. In addition, it is possible to correlate the input thermal networks analogous to an electrical circuit, representing
variable to output variable for instance ambient temperature to the building after determining values of the heat sources (internal
building load. Typically, statistical methods function best for short appliances, occupants and solar radiation), thermal resistors and
horizon predictions. Linear regression (LR) is a simple statistical capacitors (building envelope elements).
approach that attempts to correlate an output variable with inputs When forecasting is concerned with the weather rather than
through multiple linear relationships of appropriately weighed building thermal behaviour, it is often referred to as numerical
coefficients. However, most commonly forecasting techniques are weather predictions (NWP). Numerical methods rely on the
based on the Box–Jenkins methodology [9–11] and combine an analysis of the evolution of a set of variables in the atmosphere
590 D. Lazos et al. / Renewable and Sustainable Energy Reviews 39 (2014) 587–603

Table 1
Comparison of forecasting techniques.

Forecasting family Method Features Major limitations Example studies

Statistical Autoregressive moving Simple, fast, relatively high accuracy, Historical data are needed, weak [16–28]
average models (ARMA and ability to account for seasonalities to in modelling non-linear patterns
ARIMA) some extent, short forecasting horizons
Autoregressive models As ARMA, with enhanced ability to As ARMA, plus require availability [29–34]
with exogenous inputs account for recent exogenous changes of exogenous variable monitoring
(ARX)
Linear regression (LR) Simple, fast, fair accuracy Weighing of coefficients is challenging, [27,34–37]
weak in modelling non-linear patterns
and seasonalities

Machine learning Artificial neural networks Accurate, no need for supervision, able Reliance on historical data, [26–28,30,38–63]
(NN) to model non-linear patterns, high computationally complex
running speed
Support vector machines Accurate, more solid architecture than Computationally complex, low running [16,64–69]
(SVM) NN, able to model non-linear patterns, speed
needs less training data than NN

Physical and numerical Engineering white and grey Highly accurate, do not rely on historical Multiple inputs needed, can be complex [8,70–90]
box methods data, physical interpretation and slow running speed
Numerical weather Highly accurate, do not rely on historical Uncertainty present, resource intensive [91–95]
prediction (NWP) and data, physical interpretation, long and time consuming, low temporal
physical weather forecasts forecasting horizons resolutions if external data are used
using external inputs

in a multidimensional calculation space, governed by equations of variables. Solar and wind power generation, HVAC load patterns
the thermodynamics, fluid dynamics and chemical reactions of the and to a certain extent lighting load and occupancy habits are
constituents of air. They function best for horizons of some hours predominantly affected by the ambient temperature, humidity,
to days ahead. Obtaining consistently accurate weather forecasts incident solar radiation, cloud formation and sometimes wind [17].
of high temporal and spatial resolution via numerical methods is Weather is perceived as the result of numerous interactions
relatively impractical and computationally complex for commer- between the atmosphere and the surface of the Earth and since
cial buildings. As such, until recently NWP applications for it is difficult to obtain information about the state of every particle
commercial building energy management have not been exten- participating in those interactions, typically data assimilation is
sively researched. conducted via spatially and temporally scattered observations [18].
Forecasting algorithms have evolved in the last two decades to Forecasts about weather variables can be acquired via statistical or
a point where they can predict the future state of the relevant physical means. In this section a range of weather forecasting
variables with high degrees of accuracy for a range of applications applications for energy management purposes are reviewed.
in building energy management. While different metrics are being An early attempt to predict ambient temperature with an
used for the evaluation of each algorithm depending on the ARIMA model for purposes of cooling load forecasting has been
application, prediction errors are generally within acceptable developed [19]. An ARMA model was decomposed in a stochastic
ranges. part and a deterministic EW model to predict ambient tempera-
Statistical approaches are simple in terms of implementation, ture and its accuracy was reported to improve by almost 10% [20].
demonstrate high simulation speeds and require low computa- A similar method of conducting temperature forecasts for load
tional power. However they are challenged when trying to inter- prediction in buildings via a combination of deterministic part and
pret non-linear relationships and rely heavily on consistent a stochastic ARMA part has also been implemented [21]. The
historical data. Machine learning techniques are considered as an technique was extended to predict humidity and solar radiation
alternative able to capture non-linear relationships without man- using a Fourier time series for the deterministic part [22].
ual estimation of the parameters. Nonetheless, the complexity of More often predicting the internal temperature, rather than
such algorithms is high and the reliance on reliable and large ambient temperature is of interest for building energy optimisa-
amounts of archived data is still present. In addition, unique tion. Mustafaraj et al. [23] modelled and compared a range of
problems such as overfitting to the training datasets may arise. statistical models with or without external inputs for predicting
Physical forecasting methods attempt to analyse the underlying internal temperature and humidity, and it was concluded that the
principles that govern the system, rather than trying to “guess” the ARMA outperformed the ARX model for temperature, but in order
input–output relationships. Parameterisation of physical models to produce accurate humidity forecasts higher complexity AR
poses a challenge and in many cases estimations have to be made models were required. ARX models were used with a range of
according to the availability of inputs and required complexity. external weather variable inputs to predict the temperatures of
A summary of the features of the different groups of forecasting monitored classrooms in a university building [24]. It was reported
approaches and the most commonly used examples found in the that only recent external variable observations of no longer than
literature related to commercial building energy management can 15 min affected the temperature, otherwise a simple AR model is
be seen in Table 1. sufficient.
In addition to temperature, solar radiation is often considered
in relevant research and was forecasted using historical data and
3. Weather variable prediction validated for modelled buildings [25]. Furthermore, local solar
radiation was predicted via a deterministic EW technique [20].
Many of the operational traits of the buildings and behavioural A different approach through simulation of the atmospheric state
aspects of occupants are inherently correlated with weather in MATLAB was proposed [26]. A multiple LR technique for short
 D. Lazos et al. / Renewable and Sustainable Energy Reviews 39 (2014) 587–603 591

term solar radiation forecasting for buildings was implemented by the statistical model, especially for horizons longer than 8 h. It can
Zhang and Hanby [27] using a combination of onsite observations be seen in Fig. 5 that even though the statistical approach was able
and third party weather forecasts. to predict the temperature trends, the fact that it deviates notably
Approaches based on NN of different architectures have been from the actual observed values undermines its applicability to
carried out to predict ambient temperature patterns for short term building management algorithms [18]. On the other hand a NWP
horizons [28] or internal temperature in selected zones of com- ensemble was able to add value to the forecasts by minimising the
mercial buildings [29–32]. Integrated multivariate weather fore- error consistently for the 5 day horizon. The thermal mass of the
casting modules based on NN have also been developed [33,34]. building responds in long time frames, thus NWP methods are
It was found that higher accuracy is obtained when multiple zone indeed superior for these horizons. Similar conclusions confirm the
interactions are considered in internal temperature prediction [35]. superiority of NWP to statistical time series models, as the standard
Another type of NN was developed by Ferreira et al. [36] and error was reduced notably for temperature, radiation and wind
included in a sensor that directly obtains and analyses data of predictions; however statistical models were proven effective for
temperature, solar radiation and cloud coverage. Common elements short term horizons [40].
of these algorithms include the necessity for onsite weather Rather than committing resources to generate weather fore-
observations and an indexing method associating these values to casts, Kwak et al. [41] used readily available weather forecasts
particular times, days and seasons. The use of NN over statistical from the Korean Meteorological Administration and developed
methods is preferred due to their ability to capture non-linear their own numerical model for solar irradiation on a case study
patterns in the evolution of weather variables [37]. In another study office building. This allowed for higher temporal resolution and
the operative temperature inside a building was found to be more flexible predictions. In a similar manner, raw real time
predicted 9% more accurately with NN than ARX [38]. Improved weather observations from the Hong Kong Observatory were
performance has been observed when NN were hybridised with obtained and fed into modules able to process and generate more
statistical approaches, such as ARMA to forecast solar radiation after relevant information for the building, such as solar heat gains,
removing non-stationarity from the time series [39]. It should be relative humidity and localised temperature evolutions [42].
noted however, that the added complexity and costs of NN External weather forecasts have been also combined with onsite
compared to simpler statistical approaches are often not justified observations to produce temperature and humidity forecasts for
by the boosts in forecasting performance [37]. building control purposes [43].
The uncertainty related to weather variables imposes a chal- In addition to hours or days ahead horizons, there have also
lenge to predictions of longer horizons. Furthermore, certain been studies that attempt to predict the long term effects of
events such as the formation of clouds or gusts of wind are not climate changes in the energy consumption of commercial build-
easily predicted via purely statistical approaches. However, such ings [44]. Van Paassen and Luo [45] developed a statistical weather
phenomena can be modelled with higher accuracy via NWP generator model that can be applied towards that end. Another
models. For instance, outputs from a NWP model (Weather statistical approach was based on generating weather simulations
Research and Forecast Model) were produced and used to develop from existing data and utilising them for long term load predic-
an ensemble of forecasts in order to account for the sensitivity of tions [46].
weather to the initial conditions and their inherent uncertainty [18]. Weather variable forecasting and especially temperature fore-
It was argued that the NWP is more accurate and thus valuable than casting forms the foundation layer of an energy management

Fig. 5. Comparison of performance of a statistical Gaussian prediction model (top) versus a NWP model (bottom) in 5-day horizon temperature forecasts [18].
592 D. Lazos et al. / Renewable and Sustainable Energy Reviews 39 (2014) 587–603

Fig. 6. Summary of key research topics discussed in existing literature in the field of weather variable prediction for building energy management.

framework, yet it is inherently difficult to minimise uncertainty In terms of solar power generation, which is the renewable
errors. Any forecasting errors propagate to higher level predictions energy form with the highest penetration in commercial buildings,
and are possibly magnified as weather variables function as inputs there is a variety of forecasting software applications that can
for most of the techniques discussed in the next sections. Due to provide accurate estimates for generation over certain periods of
constraints on data and resources, weather forecasting for pur- time. Based on models that take into account solar radiation data,
poses of energy management in commercial buildings is com- system specs and efficiencies, these methods are very effective in
monly obtained from third parties. However, localised forecasting predictions of the system's overall performance. However, they fail
as described in this section offers a range of benefits: the ability to to capture the solar generation in real time or short horizons, thus
better capture the microclimate of the building, tailored and they have limited uses for smart-metering or DR purposes [48].
outputs of data that can be fed directly into integrated optimisa- Forecasts of real-time and short term power output of photo-
tion frameworks and customised temporal resolutions. NWP voltaic (PV) systems are much more valuable for commercial
models provide the highest capacity to minimise uncertainty, building energy management. Recurrent cloud formation patterns
especially for horizons of some hours up to days ahead. For shorter have been recognised as being notoriously difficult to predict and
horizons that are of significance for DR measures in a commercial since the power output of PV system is directly proportional to the
setting, statistical or NN approaches with external outputs seem to incident sunlight, this poses a challenge to forecasting attempts.
be superior. External outputs allow assigning heavier weights to Such a stochastic ARIMA model for cloud coverage was partially
recent observations of the weather variable in the time series, successful [49].
taking into account the trend of weather conditions to persist and Exogenous weather inputs have been reported to improve the
evolve smoothly over time [47]. As such, there is basis to argue accuracy of generation forecasts. An ARX model with NWP
that integrated hybrid statistical and numerical weather forecast- temperature inputs for online solar PV generation prediction was
ing models can function as a flexible technique to provide weather proposed by Bacher et al. [47] and demonstrated significantly
variable inputs for energy management in commercial buildings. improved accuracies by 35% compared to the persistence model.
With this kind of tailored information in hand, predicting the The model was expanded to incorporate NWP inputs for online
generation from RES is significantly simplified. Fig. 6 summarises solar thermal generation predictions with similarly high accuracy [50].
the main points examined by the literature in the field of weather For both cases, it was reported that NWP were most useful in horizons
variable forecasting for use within building energy management. longer than 4 h up to several days ahead, while the AR part was
sufficient for short term generation forecasting. This is in alignment
with the findings from the weather variable prediction section. ARX
4. Generation forecasting methods has been used in a different study and improved the AR reference
model by 13% [51]. A LR method for long horizon PV power generation
The penetration of DG, such as solar panels, wind turbines, has been carried out with ambient temperature and solar radiation
cogeneration, fuel cells or other types of batteries in commercial as exogenous inputs and reported mean monthly errors of as low as
buildings introduces a new challenge in forecasting. In the case of 5% [52]. A range of NWP outputs were used in [53] as input to a NN
cogeneration and batteries, the generation can be adjusted at will for the real-time prediction of PV solar power and produced average
and therefore optimisation algorithms are mainly concentrating forecast accuracies of 90%. A NN approach to real-time power
on the forecasts of energy costs. However, with intermittent generation forecasting has been implemented with weather data as
energy sources such as solar radiation or wind power, forecasting the inputs and it was shown that it could generate forecasts of up to
the availability of energy constitutes a major aspect of commercial 30 min ahead with notable accuracy within the 95% confidence
building energy management. intervals [54]. Furthermore, a range of statistical models with weather
 D. Lazos et al. / Renewable and Sustainable Energy Reviews 39 (2014) 587–603 593

Fig. 7. Summary of key research topics discussed in existing literature in the field of DG forecasting in commercial buildings.

inputs for generation prediction were compared and it was found that the accuracy of load prediction for commercial building energy
the genetic algorithm technique, based on evolving a set of solutions management.
in the objective space outperformed the rest [55]. Statistical methods have been broadly used in the literature for
Instead of using NWP outputs, third party weather observa- load forecasts. There is a multitude of statistical approaches
tions were obtained and used for the prediction of solar genera- depending on the application, required accuracy and resolution,
tion via an ARX framework. It was shown that it produced 15% however the typical approach is to use archived data to generate a
higher accuracy than a non-adaptive NN [56]. Additionally, NN of function that fits the load data as accurately as possible. For the
different architectures with no weather inputs at all have been energy management of commercial buildings short term load
designed and used for power forecasting [57]. forecasting (STLF) is typically of higher priority. Generally speak-
The constraint of an algorithm that is easy to use, fast and not ing, the variables associated with weather and load are changing
demanding in computational resources applies in energy avail- smoothly and slowly over short periods of time. Using historical
ability forecasts for commercial buildings. Di Piazza [58] showed load data to develop statistical algorithms is a common practice
that using a statistical k-clustering algorithm it is possible to post in order to produce forecasts for horizons up to 6 h ahead [59].
process raw data from small scale wind and solar installations and As Mathieu et al. [2] showed, demand response (DR) measures
achieve relatively high accuracy with as low as 50% of the implicate significant potential for savings if applied timely. Ambi-
original data. ent dry bulb temperature is recognised as the primary contribut-
Solar based generation (PV or solar thermal) is by far the main ing variable to commercial building loads [60], now a commonly
topic of generation forecasting literature. While wind energy has accepted fact in load prediction related research.
limited contribution in the DG mix and non-RE DG sources are not Electrical and heat loads have been also predicted using LR and
strongly affected by weather variables, they are still discussed in generalised long-term profiles for different types of buildings [61].
terms of optimisation of the energy management in commercial Fernandez et al. [62] tested a prediction algorithm with a poly-
buildings. An outline of the findings of the reviewed literature can nomial of varying degree with hourly data from a university
be found in Fig. 7. Information about the availability of energy in building in Spain, but found it can be improved if AR models were
the future from any DG sources augments the value of the energy applied.
generated, since it can assist with peak shaving, load scheduling A standard method for modelling the load time series for STLF
and DR implementations. Hence, the savings potential from is based on EW techniques. The original HW model is based on
generation in buildings increases with accurate forecasts. The three components: level, trend and seasonality [12]. The expo-
main variables associated with solar generation (solar radiation) nential weighting can be complemented by a moving average
and wind generation (wind speed) are not primarily favoured in (MA) part as well as ambient temperature dependencies [63].
load forecasts and consequently there is a conflict for resources. Taylor applied double and triple exponential smoothing in order to
Additionally, since not all buildings are designed with DG capacity, capture additional seasonal patterns in electrical loads [59,64,65],
the integration of such forecasts in existing building management however the algorithms have not been thoroughly tested on case
systems may prove too costly and complex. study buildings. Towards the opposite direction of simplifying the
model by modifying the level and removing the trend component,
He and Zhang [66] validated their approach by forecasting the AC
5. Load forecasting load of an office building.
An alternative statistical method based on historical data that is
In most commercial buildings, historical load data is readily undeniably more popular in building STLF is the AR family.
available. As a result, load forecasting for both short and long Predictions of HVAC loads for commercial buildings based on
horizons is favoured over weather or generation forecasts. The ARIMA algorithms have also been applied [67]. Fernandez et al.
simplest models treat load as a time series by itself, without any tested an ARIMA model [62] and then improved it by introducing a
correlation to weather outputs. However, it is common to use variable learning window mechanism that accounts for days of the
externally or locally generated weather inputs in order to enhance same type (weekdays, weekends) and using weight factors to
594 D. Lazos et al. / Renewable and Sustainable Energy Reviews 39 (2014) 587–603

magnify the importance of the most recent observations in the superior performance in correlating weather and time inputs with
time series [68]. Hybridisation of AR with other statistical or load outputs. Examples of networks include simple back-propaga-
machine learning methods is not uncommon in the literature. tion NN [76,77], multiple perceptron architecture [78–80], general
Xuemei et al. report that the errors of an ARIMA algorithm were regression architecture [81], hybridisation with rough
reduced by roughly 50% by post-processing the outputs with a sets [82], real-time adaptive NN with dynamic structure [83],
SVM model in predicting the cooling loads of a commercial recurrent NN emphasising on load dependencies on time [84] and
building [16]. The improved performance of the hybrid statistical NN with global solutions [74]. An algorithm based on NN has also
and SVM model over both its individual components can be seen been used to predict the occupancy heat generation, which in turns
in Fig. 8. Similar results were achieved by the authors in an effort affects the building's HVAC loads [85]. The methods display a range
to hybridise an SVM and a genetic algorithm model as the error of differences in the approaches taken to acquisition of data and
was reduced by almost 15% [69]. A hybrid model with LR and ARX pre-processing, training, weighting and post-processing. Neverthe-
has been also implemented using indices and succeeded in less, the similarity among most NN is the need for extensive
minimising the size and complexity of the dataset [70]. Gould historical data and specificity to the training data, but in a number
et al. [1] expanded the established approaches of forecasting via of comparative studies the load prediction accuracy is notably
ARIMA or EW by developing a multiple seasonality model with the higher than statistical methods [62,86,87]. Nevertheless, NN have
extra capacity to capture seasonal cycles in the time series. The been used for HVAC load forecasting in combination with ARIMA
algorithm was tested on utility data, however it could potentially and LR models after being weighed accordingly using an hierarch-
apply to commercial loads as well. ical process [88].
Another interesting statistical approach was that of Frank and Techniques based on SVM are often treated as superior to NN,
Sen [71], in which peak and overall loads of certain buildings were due to their advantages discussed in Section 2. The parameterisa-
calculated based on an algorithm that manipulates data in Energy tion of the SVM model impacts its accuracy and has been done
Databases, rather than utilising on site measurements. Finally, a via swarm particle optimisation algorithms [89,90] or data cluster-
more complex, yet effective method that is able to better capture ing [91]. In their papers, Fernandez et al. [62,87] concluded that a
the non-linear correlations between temperature and commercial 10 dimensional SVM model clearly outperformed a NN model with
building load based on Fourier series has been discussed in the a hidden layer of 10 neurons in terms of forecasting accuracy.
literature [72,73]. In addition the NN design and parameterisation poses a greater
An increasingly popular group of load forecasting methods challenge. The improved performance of SVM over NN in predict-
involve the use of NN. Network parameterisation is highly variable ing commercial loads has been also confirmed in other studies
and depends on the architecture proposed in each study. The [92,93].
energy demand is expressed as a time series depending on a Physical methods often find application in HVAC load forecast-
variety of inputs. Regarding weather components, most often the ing of commercial buildings and are commonly part of a broader
inputs that are considered are temperature or humidity [74–76]. optimisation framework. Powerful simulation software suites that
An example of the framework of a NN model utilising both are based on physical methods include Energy Plus, DOE-2 and
archived weather observations and rolling weather forecasts is TRNSYS [94–96], but lightweight computer modelling algorithms
shown in Fig. 9. Regarding commercial building load, there is no have been proposed as well, [97,98]. Energy simulation models
consensus as several algorithms have been tried and demonstrated have been used to examine commercial buildings' energy profile
from various perspectives and to different degrees of detail, thanks
to their flexibility and modularity. The models are of interest to
researchers, building designers and energy managers and are able
to handle archived and real-time load and weather data. They are
often used to both enhance the understanding of the building's
energy profile and heat transfers as well as offer valuable insight
for the design of future buildings [99]. Prediction accuracies using
energy simulations have been reported to be rather good in certain
Fig. 8. Comparison of accuracy of SARIMA, SVM and their hybrid model for the cases, as high as 99% for the prediction of HVAC loads [100,101];
prediction of cooling load of a commercial building in China [16]. however, there are also limitations for instance when predicting
solar heat gains [102]. There are additional inherent challenges
associated to complex occupant behaviour and the poor evaluation
of equipment performance. Nevertheless, in a study by Chua and
Chou [103] the simulations managed to accurately evaluate the
dependence of HVAC load to certain weather parameters. Com-
bined with statistical methods, energy simulation models have
been also successful in generating accurate predictions of load in
mixed purpose commercial buildings [104]. Another successful
application that is useful for implementing optimised controllers
was the prediction of the effects of using blinds for shading via
building energy models on building load [105,106].
The parameterisation of the models for conduction, convention
and radiation heat transfers has been thoroughly described [107].
The American Society of Heating, Refrigerating and Air Condition-
ing (ASHRAE) method uses a set of transfer functions parame-
terised appropriately to convert heat gains into cooling loads and
has been featured as the starting point in various research papers
for the estimation of commercial building load [108,109]. Rabl and
Fig. 9. Steam load prediction framework using a NN with inputs of archived load Norford [110] developed an algorithm using physical state equa-
and weather data as well as weather forecasts [79]. tions to describe the thermal behaviour and load of a case study
 D. Lazos et al. / Renewable and Sustainable Energy Reviews 39 (2014) 587–603 595

office building, while recent research [111–113] achieved the same The states at different locations within a building can be
goal through training their model with field data retrieved from expressed as a function of weather inputs, internal heat gains
various onsite data or external weather data from third parties and structural parameters, which allows the forecast of the future
[114]. A common modelling technique is based on using an thermal state and thus the energy demand [118]. Similar techni-
electrical circuitry analogy to analyse the thermal behaviour of ques, where the evolutions of the state and demand of a zone
different zones within a building. In the example of Fig. 10. within a commercial building are estimated via thermal networks
temperature differences are equivalent to voltage differences, [119] or thermodynamical and fluid mechanics equations [120]
resistors and capacitors represent the thermal resistance and have been proposed. A simplified, yet efficient approach to pre-
thermal storage capacity of the materials of the building and heat diction of cooling load was proposed by Sun et al. [94]; rather than
sources are modelled as current sources. just modelling the thermal performance of a building, a three step
An alternative approach was studied by Luo and Ariyur [115], algorithm taking into consideration seasonal and weather inputs
who included dynamic parameterisation and examined the con- as well as historical data is implemented and validated in a large
nectivity between different zones within the same building. commercial building in Hong Kong.
Variations in internal heat gains, potentially caused by changes Rather than analysing a commercial building as a single entity,
in occupant numbers or stochastic patterns, like leaving doors or Escrivá et al. [121] proposed energy auditing and performance
windows open were shown to have a significant effect in the HVAC modelling of individual air conditioning units within a building,
load. The parameterisation of the building thermal network has the aggregation of which allows for the prediction of the total load.
also been carried out with the help of a genetic algorithm [116]. Physical models can also assist when predictions of the overall
Using an alternative approach, a non-parametric state model able load, rather than the peak load are needed. An easy to implement
to predict the thermal load of commercial buildings has also been algorithm uses temperature variations to predict cooling (or
developed [117]. heating) degree hours and in turn the HVAC load for a building
[122].
Regardless of the approach, load predictions serve as inputs for
manual and automated decision making in energy management.
For STLF, many of the papers reviewed in this section report low
errors and superior prediction accuracy. Nevertheless, it shall be
noted that for statistical and machine learning methods, the
training is implemented using historical data from the case study
building. This gives rise to two issues: the ability to adapt to
changes in the building is limited, especially for statistical meth-
ods and it is not guaranteed that the algorithm can provide
consistent results if applied to a different building. Physical models
on the other hand are based on established physical laws and
equations, which are universal; thus the main hurdle for forecast-
ing accuracy is posed by resource limitations when modelling the
system. Hybridisation of models could enhance accuracy, but this
raises the complexity. As discussed earlier, locally generated
external inputs for load forecasts when available, such as tem-
perature or solar heat gains, also result in improved performance.
Consequently, it is suggested that since all families of approaches
for STLF demonstrate comparable errors and possess unique
Fig. 10. An RC thermal model used for analysing the thermal response and advantages and disadvantages, the defining factor for utilisation
forecasting HVAC load in a commercial building in Iowa, USA [113]. by a commercial building should be the ability to integrate with an

Fig. 11. Summary of key research topics discussed in existing literature in the field of commercial building load forecasting.
596 D. Lazos et al. / Renewable and Sustainable Energy Reviews 39 (2014) 587–603

optimisation framework. Depending on the size, occupancy pat- meteorological entities. However, several attempts indicate super-
terns, location and stakeholder intentions the chosen optimisation ior optimisation results and increased savings when tailored
framework should dictate the most appropriate architecture for forecasts of weather and indoor temperature obtained from onsite
STLF. The key issues discussed in the relevant literature reviewed simulations are used as inputs for MPC [25,42,75,118,127–129].
above are presented in Fig. 11. Zavala et al. [18,40] proposed a modified version of a dynamic RTO
framework that incorporates weather forecasts of different hor-
izons and concluded after simulations in a case study large
building that higher savings up to 30% can be achieved for one
6. Energy management optimisation
day ahead horizons compared to the base reactive RTO scenario.
In fact the savings were found to be even more significant for a
Based on the information about the current and future state of
modelled building with thicker insulation due to better utilisation
weather variables, load or generation, building management
of its thermal mass. In addition, the weather component resulted
systems (BMS) can adjust the set points for HVAC, as well as
into alleviating one of the main limitations of RTO routines,
regulate the flow and consumption of energy within the building
namely the inability to utilise existing trends in the time series.
or the exchange of energy with the grid. This allows the system to
It can be seen in Fig. 13, that the statistical forecasts resulted in
meet DR measures and reduce peak and annual energy consump-
higher error and even in setpoint temperatures outside the
tion and hence minimise energy costs [123]. Stochastic optimisa-
comfort zone. On the other hand, NWP predictions helped max-
tion processes have been proposed and validated with models or
imise the savings by matching predicted and actual temperatures
site testing in the literature and will be reviewed in this section.
and correctly adjust HVAC within the comfort zone. Occupancy
Sets of rules in the form of controllers regulate the parameters
comfort is regarded as an equally important objective to energy
of operation of HVAC as well as other components of the load and
savings in the optimisation literature. However, as seen in Fig. 14
can integrate the thermal mass or alternative storage sources in
there is often great margin for improvement. The optimal con-
many ways. Unlike Rule Based Control, where a set of rules govern
troller with perfect weather forecasting would achieve the lowest
the behaviour of the HVAC system deterministically, Model Pre-
energy costs at the lowest possible discomfort cost. Most conven-
dictive Control (MPC) is an optimisation framework that seeks an
tional controllers though are operating without weather inputs
optimal solution to an objective function of load bound to certain
and can only achieve low occupant discomfort at relatively high
constraints (tenant comfort, energy costs and building character-
energy costs. Dynamic RTO with weather inputs are also consid-
istics). MPC is taking into consideration the future state of the
ered and have been validated via mixed integer linear programing
energy system, the capacity of thermal or electrical storage and
[130] or NN controllers [30,33,34] with reported savings in the
the interdependencies on weather or occupancy patterns [118].
vicinity of 25–30%.
Often based on real time optimisation (RTO), these strategies can
In addition to weather inputs, occupancy pattern inputs have
minimise the costs of the system without compromising the
been considered. Significant saving potentials have been realised
comfort of the occupants as a response to external stimuli. This
and it has been concluded that weather and occupancy patterns
is mainly achieved by exploiting the whole range of the thermal
are correlated, hence reducing the degree of complexity of the
comfort zone throughout the day. External stimuli, such as
optimisation problem [131]. Using a numerical analysis it has also
temperature or solar heat gains are realised in the form of
been concluded that savings of up to 50% can be achieved with
disturbances and integrated in the algorithm in a real-time
RTO systems in place receiving weather sensitive and occupancy
manner. MPC can be adjusted to optimise the system flows on
inputs [132]. It has been demonstrated in various control and
varying horizons depending on the HVAC capacity and structure of
optimisation routines that using the outputs from a physical load
the building [124]. Typically, disturbances are forecasted using
forecasting model and taking advantage of the stored energy in
methods discussed in the previous sections. MPC demonstrates
the thermal mass of a commercial building are very effective in
tolerance to the inherent inaccuracies of the weather and demand
generating energy savings for peak consumption by shifting the
forecasts that provide inputs [125]. A flowchart demonstrating a
load to lower cost off-peak zones [133–141] mainly through
typical MPC architecture as described in [118] can be seen in
preconditioning. Lee and Braun [113] studied two such alternative
Fig. 12.
preconditioning strategies to the typical deterministic night setup.
Weather forecasting for MPC mainly incorporate predictions of
It can be seen in Fig. 15 that with preconditioning the daily load
ambient temperature and sometimes humidity and solar radiation
overall is not significantly affected, however the peak load is
[126]. Such generic inputs are often obtained from external
greatly reduced thus leading to savings without compromising
the comfort of the occupants. The performance of such controllers
can be enhanced by adding real time electricity price feedback to
the input mix [142]. Besides assisting with thermal storage
strategies and load shifting, weather variable forecasting has been
used to adjust solar heat gains via shading controls [143].
An integrated active and passive cooling, ventilation and
shading plan has been simulated with the aid of a MPC and
achieved significant energy savings of up to 80% compared to the
manual control scenario [144]. Additionally, the comfort of occu-
pants was not violated during the simulation period. Dynamic
optimisation of HVAC setpoints has also been the subject of a
similar algorithm [145]. Moreover, a single zone MPC has been
proposed using the outputs from a NN [146]. A design able to
generate a real-time model of the case study building's thermal
performance and then automate the BMS optimisation processes
has been developed [147]. In the controller discussed in another
study solar heat gains were modelled thoroughly for the energy
Fig. 12. Example of a MPC framework incorporating weather predictions [118]. management of a passive solar commercial building [148].
 D. Lazos et al. / Renewable and Sustainable Energy Reviews 39 (2014) 587–603 597

Fig. 13. Comparison of 5-day ahead operating HVAC strategies, with predicted interior temperatures being in grey, actual observations in dashed blue lines and the comfort
zone in thick solid lines [18]. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 14. Correlation of discomfort cost (Jd) and energy cost (Je) for a range of
controllers [148].

Certain studies are concerned with the optimisation of indivi-

dual components of the HVAC load, rather than developing an
integrated plan. In a consolidated study [149] a variety of strate-
gies of operation were tested for each component of the cooling
system in order to realise the optimal mode. Sun et al. [150]
proposed an optimisation algorithm to determine the most effec- Fig. 15. Effects of deterministic approach versus preconditioning on interior
tive schedule in starting up chillers in case study office buildings. temperature and daily cooling loads [113].

Other optimisation studies are delving into the control of shading

devices, such as blinds and reported savings of up to 50% from the optimisation processes have been proposed using mixed integer
base scenario [105,106]. linear programming [151]. A network flow model based on the
Optimisation of DG schedules is equally important to demand physical operation modes of a cooling, heating and power system
consumption optimisation. Regarding cogeneration systems able (CHP) as well as dynamic states of the grid has been described and
to produce both heat and electricity in commercial buildings, utilised for savings optimisation [152], while the same goal was
598 D. Lazos et al. / Renewable and Sustainable Energy Reviews 39 (2014) 587–603

Fig. 16. Summary of key research topics discussed in existing literature in the field of commercial building energy optimisation frameworks.

reached with an even more simplified flow approach [153]. buildings with both active and passive energy storage have been
Weather conditions responsible for transient trends in a CHP demonstrated [160].
system have been considered and allowed the development of a Ice storage is an alternative to utilising the thermal mass of a
predictive control algorithm [154]. building. The operating principle of ice storage system is cooling
When renewable energy sources are part of the DG mix, water down to form ice during low energy cost times and
optimisation problems become more complex due to the inter- releasing its latent heat during peak times to provide cooling.
mittent nature of power generation. The controller in addition to As with thermal storage, it is possible to minimise the costs with
optimising load variables according to the predicted states of the accurate load predictions. MPC algorithms with distinct emphasis
system, has to be able to forecast the generation. Since PV on ice storage have been described [75,137,161,162]. Typically, ice
generation is strongly affected by weather parameters and espe- storage optimisation is conducted on a daily horizon, however it
cially solar radiation and ambient temperature, such controllers was shown that control based on calculations of the load and
are relying on accurate weather forecasts. For instance, a simple weather evolution over a weekly period are viable [163].
MPC was developed and fed weather data from a local station as While there is an extensive array of issues of interest to
well as real-time pricing of energy to optimise the energy flows of research, the literature is mainly concerned about the develop-
a smart research building in Denmark [155]. Optimisation with a ment of controllers able to receive dynamic feedback and weather
controller able to monitor the online solar PV generation has also variable inputs. Active and especially passive storage via the
been developed [156]. Simple controllers that disregard weather thermal mass of the building are discussed at length. Most modern
inputs can still improve the utilisation of solar PV and thermal attempts for energy management optimisation are considering in
systems as seen in a small building case study [157]. detail the trade-off between energy savings and occupant comfort
Very often, optimisation solely involves electrical or thermal as opposed to earlier controllers focusing exclusively to energy
storage in a commercial building. A non-predictive approach has savings. A summary of the above findings can be seen in Fig. 16.
been developed, in which energy storage charge and discharge
rates are governed by the feedback generated from previous
responses of the system in terms of costs [158]. Batteries can be 7. Discussion
regarded as additional nodes of DG in a building, whether
stationary or mobile. Electric vehicle batteries connectivity to a Direct comparison of the reported accuracies in prediction of
commercial building have been modelled using a mixed integer weather variables and energy consumption as well as the observed
linear programme in order to predict and optimise the behaviour savings via optimisation algorithms is not particularly meaningful,
of the energy system [159]. In addition, the effects of MPC on as most of the techniques are relevant and tailored to specific
 D. Lazos et al. / Renewable and Sustainable Energy Reviews 39 (2014) 587–603 599

buildings and depend on the availability of data and required optimal results for the majority of aspects of energy management
outputs. Hence, it is not simple to establish an absolute superiority in certain commercial buildings. The framework of the proposed
of a specific group of techniques for the purpose of energy integrated architecture consisting of a three layer forecasting and
management in commercial buildings. Instead and as seen earlier, optimisation system can be seen in Fig. 17. The interdependencies
there are distinct advantages and disadvantages depending on the between modules of successive layers have been established in the
available resources, complexity and magnitude of each case where literature and reviewed in the sections above. A framework such
a forecasting algorithm is used to assist in energy management. as the one shown in Fig. 17 is meaningful for commercial buildings
After reviewing the literature however, it is evident that with enhanced energy generation and management capacity, so
ambient temperature, as well as humidity and solar radiation are that the potential added savings can justify the increased
the main factors of uncertainty in building energy consumption model cost.
and generation. As such, the major finding from the analysis above The main advantage of this modular approach is the added
is that the consideration of weather variable forecasting always value in a range of building energy management components of
produces enhanced accuracy compared to a deterministic the higher tier. The uninterrupted flow of inputs and outputs from
approach of prediction and optimisation of the system with non- each layer to the next, the compatibility of data formatting and
weather sensitive data. tailored control over all aspects of energy management are also
A common technique that was seen in many case studies, noteworthy benefits. Since the uncertainty associated with
involves predictions of temperature, which are then in turn weather variables propagates towards the higher levels, it is
applied to HVAC load forecasts and management of the set points important to be able to have access to reliable forecasts for the
or other DR measures, such as preconditioning or active storage. range of variables required for different horizons, preferably
Such one-dimensional approaches impose constraints on the localised. Weather variable predictions in level 1 feed into higher
energy savings margins. For larger commercial buildings with a level modules and assist with the respective predictions. Tem-
range of generation and management options additional weather perature is the most important variable and can be predicted using
variables are significant for the analysis of the energy system. a time series or a NN approach especially for short horizons;
However, the implementation of an energy system able to utilise however, NWP models are favourable for such an integrated
several weather inputs appears to be problematic. The introduc- framework as they can predict factors like solar radiation, humid-
tion of additional input variables increase the degree of complexity ity, wind speed and direction and cloud formation that contribute
and in turn the required computational resources. Furthermore, to the system's performance.
weather variables require non-linear models that cannot be Level 2 consists of modules related to load and availability of
accurately designed using statistical methods. The incorporation energy. Energy generation forecasts from RES are useful for
of localised numerical weather forecasts allows accurate handling managing peak and base loads as well as energy flows to and
of non-linear models and while still there is an increase in from the grid. Passive energy storage in the thermal mass of the
complexity and cost of the energy system, we argue that the building can also be modelled towards the same end. Other
savings potential is promising. important parameters depending on weather forecast outputs
So far, the literature has not considered the integration of a include solar heat gains and occupant comfort. In addition, climate
complete set of weather variable forecasting for commercial evolution prediction may be possible, which is useful for future
building energy management. Using a set of lightweight numerical planning and upgrading. At this level, electricity pricing is con-
prediction tools and statistical processing, we propose that a sidered as a major non-weather dependent component subject to
modular approach based on weather predictions is able to provide forecasting.

Fig. 17. Proposed integrated weather, load and generation forecasting architecture for energy system optimization.
600 D. Lazos et al. / Renewable and Sustainable Energy Reviews 39 (2014) 587–603

Level 3 consists of the energy management optimisation [8] Sun Y, Wang S, Xiao F, Gao D. Peak load shifting control using different cold
framework, typically a holistic MPC. Besides HVAC operation thermal energy storage facilities in commercial buildings: a review. Energy
Convers Manag 2013;71:101–14.
optimisation, active storage where available such as batteries or [9] Box GEP, Jenkins G. Time series analysis, forecasting and control. San
ice storage, is relying on inputs from the previous layer. Electricity Francisco, CA, USA: Holden-Day; 1970.
flows within the building (from DG sources to load and/or storage) [10] Hagan MT, Behr SM. The time series approach to short term load forecasting.
IEEE Trans Power Syst 1987;PWRS-2:785–91.
as well as exchange with the grid can be planned accordingly. [11] Loveday DL, Craggs C. Stochastic modelling of temperatures for a full-scale
External factors that need to be considered include the building occupied building zone subject to natural random influences. Appl Energy
type, size, equipment and occupancy patterns. The cost of the 1993;45:295–312.
[12] Wan Ahmad WKA, Ahmad S. Arima model and exponential smoothing
complexity of the system and higher spatial and temporal resolu-
method: a comparison. In: Proceedings of the 20th national symposium on
tions should always be weighed against the potential savings in mathematical sciences: research in mathematical sciences: a catalyst for
each layer. creativity and innovation. USA: American Institute of Physics; 2013. p. 1312–21.
[13] Holt CC. Forecasting seasonals and trends by exponentially weighted moving
averages. Int J Forecast 2004;20:5–10.
[14] Taylor JW. Short-term electricity demand forecasting using double seasonal
8. Conclusion exponential smoothing. J Oper Res Soc 2003;54:799–805.
[15] Rumelhart DE, Hinton GE, Williams RJ. Learning representations by back-
propagating errors. Nature 1986;323:533–6.
The aim of this review was to investigate the multiple dimen- [16] Xuemei L, Lixing D, Yuyuan D, Lanlan L. Hybrid support vector machine and
sions and value of forecasting and energy optimisation algorithms ARIMA model in building cooling prediction. In: Proceedings of the 2010
for commercial building energy systems, with emphasis on those international symposium on computer, communication, control and auto-
mation, 3CA 2010. Tainan, Taiwan: IEEE Computer Society; 2010. p. 533–6.
with weather variable inputs. While annual or longer horizon [17] Pedersen L. Use of different methodologies for thermal load and energy
forecasting is certainly valuable for planning upgrades and poli- estimations in buildings including meteorological and sociological input
cies, most building management systems are focusing on short parameters. Renew Sustain Energy Rev 2007;11:998–1007.
[18] Zavala VM, Constantinescu EM, Anitescu M. Economic impacts of advanced
term load forecasting and high resolution dynamic optimisation.
weather forecasting on energy system operations. In: Proceedings of the
A review of forecasting concepts was conducted and complemen- innovative smart grid technologies conference, SGT 2010. Gaithersburg, MD,
ted by a thorough discussion of relevant studies in the literature United states: IEEE Computer Society; IEEE PES – Power and Energy Society;
2010.
for weather, generation and load forecasting for commercial
[19] MacArthur J, Mathur A, Zhao J. On-line recursive estimation for load profile
buildings as well as their applications in energy management. prediction. ASHRAE Trans 1989;95:621–8.
It was found that weather variables are significant components [20] Ren MJ, Wright JA. Adaptive diurnal prediction of ambient dry-bulb
of the evolution of building energy systems and minimising the temperature and solar radiation. HVAC and R Res 2002;8:383–401.
[21] Yoshida H, Terai T. ARMA type weather model for air-conditioning, heating
uncertainty in predicting their evolution can lead to significant and cooling load calculation. Energy Build 1991;16:625–34.
savings, usually in the range of 15–30% compared to a determi- [22] Yoshida H, Terai T. Modeling of weather data by time series analysis for air-
nistic and non-weather sensitive control approach. For smaller conditioning load calculators. ASHRAE Winter Meeting, January 25, 1992–
January 29, 1992. Pt 1.. Anaheim, CA, USA: ASHRAE; 1992; 328–45.
building systems with limited optimisation components, simple [23] Mustafaraj G, Chen J, Lowry G. Development of room temperature and
statistical processing of archived data or even external forecasts of relative humidity linear parametric models for an open office using BMS
temperature are sufficient for energy management. For larger and data. Energy Build 2010;42:348–56.
[24] Rios-Moreno GJ, Trejo-Perea M, Castaneda-Miranda R, Hernandez-Guzman VM,
more dynamic building energy systems, it was found that the
Herrera-Ruiz G. Modelling temperature in intelligent buildings by means of
availability of accurate weather forecasts greatly enhances the autoregressive models. Autom Constr 2007;16:713–22.
savings potential. Commercial buildings with such capacity can [25] Chen T-Y, Athienitis AK. Ambient temperature and solar radiation prediction
for predictive control of HVAC systems and a methodology for optimal
benefit from integrated weather forecasting frameworks as there
building heating dynamic operation. ASHRAE Trans 1996;102:26–35.
are dynamic interactions between individual components of the [26] Keller B, Costa AMS. A Matlab GUI for calculating the solar radiation and
weather inputs, the system and the grid. While most approaches shading of surfaces on the earth. Comput Appl Eng Educ 2011;19:161–70.
utilise temperature forecasts, it is evident that additional factors [27] Zhang Y, Hanby VI. Short-term prediction of weather parameters using
online weather forecasts. Building Simulation 2007, BS 2007, September 3,
contribute to different extents to energy management and opti- 2007–September 6, 2007. Beijing, China: International Building Performance
misation. So far there are no practical lightweight models towards Simulation Association; 2007; 1411–6.
that end in the literature. Hence, we argue that integrated weather [28] Lanza PAG, Cosme JMZ. A short-term temperature forecaster based on a
novel radial basis functions neural network. Int J Neural Syst 2001;11:71–7.
forecasts are vital and may potentially generate additional savings, [29] Mustafaraj G, Lowry G, Chen J. Prediction of room temperature and relative
especially where generation from renewable energy sources or humidity by autoregressive linear and nonlinear neural network models for
passive thermal storage are available. an open office. Energy Build 2011;43:1452–60.
[30] Ruano AE, Crispim EM, Conceicao EZE, Lucio MMJR. Prediction of building's
temperature using neural networks models. Energy Build 2006;38:682–94.
References [31] Gouda MM, Danaher S, Underwood CP. Application of an artificial neural
network for modelling the thermal dynamics of a building's space and its
heating system. Math Comput Model Dyn Syst 2002;8:333–44.
[1] Gould PG, Koehler AB, Ord JK, Snyder RD, Hyndman RJ, Vahid-Araghi F. [32] Yang I-H, Kim K-W. Prediction of the time of room air temperature
Forecasting time series with multiple seasonal patterns. Eur J Oper Res descending for heating systems in buildings. Build Environ 2004;39:19–29.
2008;191:205–20. [33] Argiriou AA, Bellas-Velidis I, Balaras CA. Development of a neural network
[2] Mathieu JL, Price PN, Kiliccote S, Piette MA. Quantifying Changes in building heating controller for solar buildings. Neural Netw 2000;13:811–20.
electricity use, with application to demand response. IEEE Trans Smart Grid [34] Argiriou AA, Bellas-Velidis I, Kummert M, Andre P. A neural network
2011;2:507–18. controller for hydronic heating systems of solar buildings. Neural Netw
[3] Perez-Lombard L, Ortiz J, Pout C. A review on buildings energy consumption 2004;17:427–40.
information. Energy Build 2008;40:394–8. [35] Huang H, Chen L, Mohammadzaheri M, Hu E, Chen M. Multi-zone tempera-
[4] Lu N, Taylor T, Jiang W, Correia J, Leung LR, Wong PC. The temperature ture prediction in a commercial building using artificial neural network
sensitivity of the residential load and commercial building load. In: Proceed- model. In: Proceedings of the 10th IEEE international conference on control
ings of the 2009 IEEE Power and Energy Society general meeting, PES'09. and automation, ICCA 2013. Hangzhou, China: IEEE Computer Society; 2013.
Calgary, AB, Canada: IEEE Computer Society; 2009. p. 1896–901.
[5] Steinfeld J, Bruce A, Watt M. Peak load characteristics of Sydney office [36] Ferreira PM, Gomes JM, Martins IAC, Ruano AE. A neural network based
buildings and policy recommendations for peak load reduction. Energy Build intelligent predictive sensor for cloudiness, solar radiation and air tempera-
2011;43:2179–87. ture. Sensors (Switzerland) 2012;12:15750–77.
[6] Zhao H-X, Magoules F. A review on the prediction of building energy [37] Florita AR, Henze GP. Comparison of short-term weather forecasting models
consumption. Renew Sustain Energy Rev 2012;16:3586–92. for model predictive control. HVAC and R Res 2009;15:835–53.
[7] Foucquier A, Robert S, Suard F, Stephan L, Jay A. State of the art in building [38] Soleimani-Mohseni M, Thomas B, Fahlen P. Estimation of operative tem-
modelling and energy performances prediction: a review. Renew Sustain perature in buildings using artificial neural networks. Energy Build 2006;38:
Energy Rev 2013;23:272–88. 635–40.
 D. Lazos et al. / Renewable and Sustainable Energy Reviews 39 (2014) 587–603 601

[39] Ji W, Chee KC. Prediction of hourly solar radiation using a novel hybrid [64] Taylor JW. Exponentially weighted methods for forecasting intraday time
model of ARMA and TDNN. Sol Energy 2011;85:808–17. series with multiple seasonal cycles. Int J Forecast 2010;26:627–46.
[40] Zavala VM, Constantinescu EM, Krause T, Anitescu M. On-line economic [65] Taylor JW, Snyder RD. Forecasting intraday time series with multiple
optimization of energy systems using weather forecast information. J Process seasonal cycles using parsimonious seasonal exponential smoothing. Omega
Control 2009;19:1725–36. 2012;40:748–57.
[41] Kwak Y, Seo D, Jang C, Huh J-H. Feasibility study on a novel methodology for [66] He D-S, Zhang X. Analysis of air conditioning load prediction by modified
short-term real-time energy demand prediction using weather forecasting seasonal exponential smoothing model. Tongji Daxue Xuebao/J Tongji Univ
data. Energy Build 2013;57:250–60. 2005;33:1672–6.
[42] Zhou Q, Wang S, Xu X, Xiao F. A grey-box model of next-day building [67] Kimbara A, Kurosu S, Endo R, Kamimura K, Matsuba T, Yamada A. On-line
thermal load prediction for energy-efficient control. Int J Energy Res prediction for load profile of an air-conditioning system. In: Proceedings of
2008;32:1418–31. the 1995 ASHRAE annual meeting. Pt 2. San Diego, CA, USA: ASHRAE; 1995.
[43] Nagai T. A method for revising temperature and humidity prediction using p. 198–207.
additional observations and weather forecasts. Building Simulation 2007, BS [68] Borges CE, Penya YK, Fernandez I. Optimal combined short-term building
2007, September 3, 2007–September 6, 2007. Beijing, China: International load forecasting. In: Proceedings of the IEEE Power and Energy Society's
Building Performance Simulation Association; 2007; 245–52. Innovative Smart Grid Technologies Asia 2011 Conference,ISGT Asia 2011.
[44] Yu FW, Chan KT, Sit RKY. Climatic influence on the design and operation of Perth, WA, Australia: IEEE Computer Society; IEEE Power and Energy Society;
chiller systems serving office buildings in a subtropical climate. Energy Build 2011.
2012;55:500–7. [69] Li X, Ding L, Li Y, Xu G, Li J. Hybrid genetic algorithm and support vector
[45] Van Paassen AHC, Luo QX. Weather data generator to study climate change regression in cooling load prediction. In: Proceedings of the 2010 3rd
on buildings. 4th ed.Chartered Institute of Building Services Engineers; International Conference on Knowledge Discovery and Data Mining (WKDD
2002; 251–8. 2010). Los Alamitos, CA, USA: IEEE Computer Society; 2010. p. 527–31.
[46] David M, Adelard L, Garde F, Boyer H. Weather data analysis based on typical [70] Kyungtae Y, Luck R, Mago PJ, Heejin C. Building hourly thermal load
weather sequences. application: energy building simulation. In: Building prediction using an indexed ARX model. Energy Build 2012;54:225–33.
Simulation 2005–9th international IBPSA conference, BS 2005. Montreal, [71] Frank SM, Sen PK. Estimation of electricity consumption in commercial
Canada: International Building Performance Simulation Association; 2005. buildings. In: Proceedings of the 43rd North American Power Symposium, NAPS
p. 239–46. 2011. Boston, MA, United states: IEEE Computer Society; Power and Energy
[47] Bacher P, Madsen H, Nielsen HA. Online short-term solar power forecasting. Society (IEEE PES); Northeastern University College of Engineering; 2011.
Sol Energy 2009;83:1772–83. [72] Dhar A, Reddy TA, Claridge DE. Fourier series model to predict hourly heating
[48] Tamizh-Mani G, Ishioye J-P, Voropayev A, Kang Y. Photovoltaic performance and cooling energy use in commercial buildings with outdoor temperature
models: an evaluation with actual field data. Reliability of photovoltaic cells, as the only weather variable. J Sol Energy Eng, Trans ASME 1999;121:47–53.
modules, components, and systems, August 11, 2008–August 13, 2008. San [73] Dhar A, Reddy TA, Claridge DE. Generalization of the Fourier series approach
Diego, CA, United states: The International Society for Optical Engineering to model hourly energy use in commercial buildings. J Sol Energy Eng, Trans
(SPIE); 2008. ASME 1999;121:54–62.
[49] Chowdhury BH, Rahman S. Forecasting sub-hourly solar irradiance for [74] Yokoyama R, Wakui T, Satake R. Prediction of energy demands using neural
prediction of photovoltaic output. Conference record of the nineteenth IEEE
network with model identification by global optimization. Energy Convers
photovoltaic specialists conference – 1987. New Orleans, LA, USA: IEEE;
Manag 2009;50:319–27.
1987; 171–6.
[75] Ferrano FJ, Wong K-FV. Prediction of thermal storage loads using a neural
[50] Bacher P, Madsen H, Perers B. Short-term solar collector power forecasting.
network. 1990 Annual meeting of the American Society of Heating,
In: Proceedings of the 30th ISES biennial solar world congress 2011, SWC
Refrigerating and Air-Conditioning Engineers, Technical and Symposium Papers,
2011. Kassel, Germany: International Solar Energy Society; 2011. p. 4579–89.
June 10, 1990–June 13, 1990. Pt 2.. St. Louis, MO, USA: ASHRAE; 1990; 723–6.
[51] Yang C, Xie L. A novel ARX-based multi-scale spatio-temporal solar power
[76] Gonzalez PA, Zamarreno JM. Prediction of hourly energy consumption in
forecast model. In: Proceedings of the 2012 North American Power Sympo-
buildings based on a feedback artificial neural network. Energy Build
sium, NAPS 2012. Champaign, IL, United states: IEEE Computer Society; 2012.
2005;37:595–601.
[52] Huang CY, Chen HJ, Chan CC, Chou CP, Chiang CM. Thermal model based
[77] Shi L, Wang J. Application of artificial neural network to predict the hourly
power-generated prediction by using meteorological data in BIPV system. In:
cooling load of an office building. In: Proceedings of the 2009 international
Proceedings of the 1st International Conference on Smart Grid and Clean
joint conference on Computational Sciences and Optimization, CSO 2009.
Energy Technologies, ICSGCE 2011. Chengdu, China: Elsevier Ltd.; 2011.
Sanya, Hainan, China: IEEE Computer Society; 2009. p. 528–30.
p. 531–7.
[78] Kwok SSK, Yuen RKK, Lee EWM. An intelligent approach to assessing the
[53] Chen C, Duan S, Cai T, Liu B. Online 24-h solar power forecasting based on
effect of building occupancy on building cooling load prediction. Build
weather type classification using artificial neural network. Sol Energy
Environ 2011;46:1681–90.
2011;85:2856–70.
[79] Kusiak A, Li M, Zhang Z. A data-driven approach for steam load prediction in
[54] Chow SKH, Lee EWM, Li DHW. Short-term prediction of photovoltaic energy
generation by intelligent approach. Energy Build 2012;55:660–7. buildings. Appl Energy 2010;87:925–33.
[55] Wu Y-K, Chen C-R, Hsu K-W, Yeh C-M. Research on short-term photovoltaic [80] Kreider JF, Wang XA. Improved artificial neural networks for commercial
power forecasting. Mon J Taipower's Eng 2011;750:99–110. building energy use prediction. In: Proceedings of the 1992 ASME–JSES–
[56] Cai T, Duan S, Chen C. Forecasting power output for grid-connected KSES international solar energy conference Part 1 (of 2). Maui, HI, USA:
photovoltaic power system without using solar radiation measurement. In: ASME; 1992. p. 361–6.
Proceedings of the 2nd international symposium on Power Electronics for [81] Ben-Nakhi AE, Mahmoud MA. Cooling load prediction for buildings using
Distributed Generation systems, PEDG 2010. Hefei, China: IEEE Computer general regression neural networks. Energy Convers Manag 2004;45:
Society; 2010. p. 773–7. 2127–41.
[57] Al-Messabi N, Li Y, El-Amin I, Goh C. Forecasting of photovoltaic power yield [82] Hou Z, Lian Z, Yao Y, Yuan X. Cooling-load prediction by the combination of
using dynamic neural networks. In: Proceedings of the 2012 annual Inter- rough set theory and an artificial neural-network based on data-fusion
national Joint Conference on Neural Networks, IJCNN 2012, Part of the 2012 technique. Appl Energy 2006;83:1033–46.
IEEE World Congress on Computational Intelligence, WCCI 2012. Brisbane, [83] Rivard H, Jin Y, Zmeureanu R. On-line building energy prediction using
QLD, Australia: Institute of Electrical and Electronics Engineers Inc.; IEEE adaptive artificial neural networks. Energy Build 2005;37:1250–9.
Computational Intelligence Society (CIS); International Neural Network [84] Ahmed O. Prediction of building thermal load using recurrent neural
Society (INNS); 2012. network. Engineering Applications of Neural Networks. In: Proceeding of
[58] Di Piazza A, Di Piazza A, Ragusa A, Vitale G. Environmental data processing the 5th international conference on Engineering Applications of Neural
by clustering methods for energy forecast and planning. Renew Energy Networks (EANN'99). Torun, Poland: Wydawnictwo Adam Marszalek;
2011;36:1063–74. 1999. p. 99–104.
[59] Taylor JW. Short-term load forecasting with exponentially weighted meth- [85] Kwok SSK. A study of the importance of occupancy to building cooling load
ods. IEEE Trans Power Syst 2012;27:458–64. in prediction by intelligent approach. Energy Convers Manag 2011;52:
[60] Reddy TA, Claridge DE. Using synthetic data to evaluate multiple regression 2555–64.
and principal component analyses for statistical modeling of daily building [86] Kawashima M, Dorgan CE, Mitchell JW. Hourly thermal load prediction for
energy consumption. Energy Build 1994;21:35–44. the next 24 h by Arima, Ewma, LR, and an artificial neural network.
[61] Pedersen L, Stang J, Ulseth R. Load prediction method for heat and electricity In: Proceedings of the 1995 ASHRAE annual meeting. 1st ed. Chicago, IL,
demand in buildings for the purpose of planning for mixed energy distribu- USA: ASHRAE; 1995. p. 186–200.
tion systems. Energy Build 2008;40:1124–34. [87] Penya YK, Borges CE, Agote D, Fernandez I. Short-term load forecasting in air-
[62] Fernandez I, Borges CE, Penya YK. Efficient building load forecasting. conditioned non-residential buildings. In: Proceedings of the 2011 IEEE
In: Proceedings of the 2011 IEEE 16th conference on Emerging Technologies International Symposium on Industrial Electronics, ISIE 2011. Gdansk,
and Factory Automation, ETFA 2011. Toulouse, France: Institute of Electrical Poland: IEEE Computer Society; 2011. p. 1359–64.
and Electronics Engineers Inc.; IEEE Industrial Electronics Society; University [88] Yao Y, Lian Z, Liu S, Hou Z. Hourly cooling load prediction by a combined
of Toulouse 3; 2011. forecasting model based on analytic hierarchy process. Int J Therm Sci
[63] Seem JE, Braun JE. Adaptive methods for real-time forecasting of building 2004;43:1107–18.
electrical demand. ASHRAE winter meeting – technical papers, January 19, [89] Xuemei L, Ming S, Lixing D, Gang X, Jibin L. Particle swarm optimization-
1991–January 23, 1991. Pt 1.. New York, NY, USA: ASHRAE; 1991; 710–21. based LS-SVM for building cooling load prediction. J Comput 2010;5:614–21.
602 D. Lazos et al. / Renewable and Sustainable Energy Reviews 39 (2014) 587–603

[90] Li X, Ding L, Li L. A novel building cooling load prediction based on SVR and [118] Oldewurtel F, Parisio A, Jones CN, Gyalistras D, Gwerder M, Stauch V, et al.
SAPSO. In: Proceedings of the 2010 international symposium on computer, Use of model predictive control and weather forecasts for energy efficient
communication, control and automation (3CA). Piscataway, NJ, USA: IEEE; building climate control. Energy Build 2012;45:15–27.
2010. p. 528–32. [119] Lehmann B, Gyalistras D, Gwerder M, Wirth K, Carl S. Intermediate complex-
[91] Cherkassky V, Roy Chowdhury S, Landenberger V, Tewari S, Bursch P. ity model for model predictive control of integrated room automation.
Prediction of electric power consumption for commercial buildings. In: Energy Build 2013;58:250–62.
Proceedings of the 2011 International Joint Conference on Neural Network, [120] Wen J, Smith TF. Development and validation of online models with
IJCNN 2011. San Jose, CA, United states: Institute of Electrical and Electronics parameter estimation for a building zone with VAV system. Energy Build
Engineers Inc.; 2011. p. 666–72. 2007;39:13–22.
[92] Li Q, Meng Q, Cai J, Yoshino H, Mochida A. Applying support vector machine [121] Escriva-Escriva G, Alvarez-Bel C, Valencia-Salazar I. Method for modelling
to predict hourly cooling load in the building. Appl Energy 2009;86: space conditioning aggregated daily load curves: application to a university
2249–56. building. Energy Build 2010;42:1275–82.
[93] Li Q, Meng Q, Cai J, Yoshino H, Mochida A. Predicting hourly cooling load in [122] Oktay Z, Coskun C, Dincer I. A new approach for predicting cooling degree-
the building: a comparison of support vector machine and different artificial hours and energy requirements in buildings. Energy 2011;36:4855–63.
neural networks. Energy Convers Manag 2009;50:90–6. [123] Luo Q, Ariyur KB, Mathur A. Real time energy management: cutting the
[94] Sun Y, Wang S, Xiao F. Development and validation of a simplified online carbon footprint and energy costs via hedging, local sources and active
cooling load prediction strategy for a super high-rise building in Hong Kong. control. In: Proceedings of the 2009 ASME Dynamic Systems and Control
Energy Convers Manag 2013;68:20–7. Conference, DSCC 2009. PART A. Hollywood, CA, United states: American
[95] Chou SK, Chang WL. Development of an energy-estimating equation for large Society of Mechanical Engineers; 2010. p. 157–64.
commercial buildings. Int J Energy Res 1993;17:759–73. [124] Krarti M, Henze GP, Bell D. Planning horizon for a predictive optimal
[96] Crawley DB, Pedersen CO, Lawrie LK, Winkelmann FC. Energy plus: energy controller for thermal energy storage systems. In: Proceedings of the ASHRAE
simulation program. ASHRAE J 2000;42:49–56. Annual Meeting. Seattle, WA, USA: ASHRAE; 1999. PART 2.
[97] Karmacharya S, Putrus G, Underwood C, Mahkamov K. Thermal modelling of [125] Henze GP, Krarti M. Impact of forecasting uncertainty on the performance of
the building and its HVAC system using Matlab/Simulink. In: Proceedings of a predictive optimal controller for thermal energy storage systems. In:
the 2012 2nd international symposium on Environment-Friendly Energies Proceedings of the ASHRAE annual meeting. Seattle, WA, USA: ASHRAE;
and Applications (EFEA 2012). Piscataway, NJ, USA: IEEE; 2012. p. 202–6. 1999. PART 2.
[98] Turiel I, Boschen R, Seedall M, Levine M. Simplified energy analysis [126] Cooperman A, Dieckmann J, Brodrick J. Using weather data for predictive
methodology for commercial buildings. Energy Build 1984;6:67–83. control. ASHRAE J 2010;52:130–2.
[99] Cloudt CL, Gomez JD, Nishimoto TK, Shephard LE. Coupling simulation tools [127] Candanedo JA, Paradis E, Stylianou M. Building simulation weather forecast
and real-time data to improve building energy performance. In: Proceedings files for predictive control strategies. In: Proceedings of the 4th Symposium
of the 2013 IEEE green technologies conference, GREENTECH 2013. Denver, on Simulation for Architecture and Urban Design, SimAUD 2013, Part of the
CO, United states: IEEE Computer Society; 2013. p. 373–9. 2013 Spring Simulation Multiconference, SpringSim 2013. 8th ed. San Diego,
[100] Li N, Kwak J-Y, Becerik-Gerber B, Tambe M. Predicting HVAC energy CA, United states: The Society for Modeling and Simulation International;
consumption in commercial buildings using multiagent systems. In: Pro- 2013. p. 22–7.
ceedings of the 30th International Symposium on Automation and Robotics [128] Aswani A, Master N, Taneja J, Krioukov A, Culler D, Tomlin C. Energy-efficient
in Construction and Mining, ISARC 2013, Held in Conjunction with the 23rd building HVAC control using hybrid system LBMPC. PART 1; 2012. p. 496–501.
World Mining Congress. Montreal, QC, Canada: Canadian Institute of Mining, [129] Cigler J, Privara S. Subspace identification and model predictive control for
Metallurgy and Petroleum; 2013. p. 987–94. buildings. In: Proceedings of the 2010 11th International Conference on
[101] Raftery P, Keane M, Costa A. Calibrating whole building energy models: detailed
Control Automation Robotics & Vision (ICARCV 2010). Piscataway, NJ, USA:
case study using hourly measured data. Energy Build 2011;43:3666–79.
IEEE; 2010. p. 750–5.
[102] Loutzenhiser PG, Manz H, Felsmann C, Strachan PA, Frank T, Maxwell GM.
[130] Collazos A, Marechal F, Gahler C. Predictive optimal management method for
Empirical validation of models to compute solar irradiance on inclined
the control of polygeneration systems. Comput Chem Eng 2009;33:1584–92.
surfaces for building energy simulation. Sol Energy 2007;81:254–67.
[131] Dong B, Lam KP, Neuman CP. Integrated building control based on occupant
[103] Chua KJ, Chou SK. A performance-based method for energy efficiency
behavior pattern detection and local weather forecasting. In: Proceedings of
improvement of buildings. Energy Convers Manag 2011;52:1829–39.
the 12th conference of international building performance simulation
[104] Gao X, Malkawi AM, Yi YK. A new method for predicting mixed-use building
association building simulation 2011, BS 2011,. Sydney, NSW, Australia:
energy: the use of simulation to develop statistical models. In: Proceedings of
International Building Performance Simulation Association; 2011. p. 193–200.
the 13th conference of international building performance simulation
[132] Zavala VM. Real-time optimization strategies for building systems. Ind Eng
association, BS 2013. Chambery, France: International Building Performance
Chem Res 2013;52:3137–50.
Simulation Association; 2013. p. 2349–56.
[133] Henze GP, Kalz DE, Liu S, Felsmann C. Experimental analysis of model-based
[105] Kotey NA, Collins MR, Wright JL. A simplified method for calculating the
predictive optimal control for active and passive building thermal storage
effective solar optical properties of a venetian blind layer for building energy
inventory. HVAC and R Res 2005;11:189–213.
simulation. J Sol Energy Eng, Trans ASME 2009;131:0210021–9.
[134] Oldewurtel F, Jones CN, Morari M. A tractable approximation of chance
[106] Daum D, Morel N. Assessing the total energy impact of manual and
constrained stochastic MPC based on affine disturbance feedback.
optimized blind control in combination with different lighting schedules in
In: Proceedings of the 47th IEEE Conference on Decision and Control, CDC
a building simulation environment. J Build Perform Simul 2010;3:1–16.
[107] Fraisse G, Viardot C, Lafabrie O, Achard G. Development of a simplified and 2008. Cancun, Mexico: Institute of Electrical and Electronics Engineers Inc.;
accurate building model based on electrical analogy. Energy Build 2002;34: 2008. p. 4731–6.
[135] Braun JE. Reducing energy costs and peak electrical demand through optimal
1017–31.
[108] Al-Rabghi OMA, Al-Johani KM. Utilizing transfer function method for hourly control of building thermal storage. In: Proceedings of the 1990 annual
cooling load calculations. Energy Convers Manag 1997;38:319–32. meeting of the american society of heating, refrigerating and air-conditioning
[109] Chen T, Yu Z. A statistical method for selection of sequences of coincident engineers, technical and symposium papers. pt 2. St. Louis, MO, USA:
weather parameters for design cooling load calculations. Energy Convers ASHRAE; 1990. p. 876–88.
Manag 2009;50:813–21. [136] Nagai T, Reddy A, Massie D Optimization method for minimizing annual
[110] Rabl A, Norford LK. Peak load reduction by preconditioning buildings at night. energy, peak energy demand, and annual energy cost through use of building
Int J Energy Res 1991;15:781–98. thermal storage. In: Proceedings of the 2002 ASHRAE Winter Meeting.
[111] Braun JE, Montgomery KW, Chaturvedi N. Evaluating the performance of Atlantic City, NJ, United States: Amer. Soc. Heating, Ref. Air-Conditoning
building thermal mass control strategies. HVAC and R Res 2001;7:403–28. Eng. Inc.; 2002. p. 43–53.
[112] Braun JE, Chaturvedi N. An inverse gray-box model for transient building load [137] Hajiah A, Krarti M. Optimal control of building storage systems using both ice
prediction. HVAC and R Res 2002;8:73–99. storage and thermal mass – Part I: simulation environment. Langford Lane,
[113] Lee K-H, Braun JE. Reducing peak cooling loads through model-based control Kidlington, Oxford, OX5 1GB, United Kingdom: Elsevier Ltd.; 2012; 499–508.
of zone temperature setpoints. In: Proceedings of the 2007 American Control [138] Ruud MD, Mitchell JW, Klein SA. Use of building thermal mass to offset
Conference, ACC. New York, NY, United states: Institute of Electrical and cooling loads. 1990 Annual Meeting of the American Society of Heating,
Electronics Engineers Inc.; 2007. p. 5070–5. Refrigerating and Air-Conditioning Engineers, Technical and Symposium
[114] Westphal FS, Lamberts R. The use of simplified weather data to estimate Papers. pt 2. St. Louis, MO, USA: ASHRAE; 1990. p. 820–9.
thermal loads of non-residential buildings. 8th ed. Elsevier Ltd; 2004; 847–54. [139] Xu PH, Piette P, Braun MA, Peak JE. Demand reduction from pre-cooling with
[115] Luo Q, Ariyur KB. Building thermal network model and application to zone temperature reset in an office building. 2004 ACEEE summer study on
temperature regulation. In: Proceedings of the 2010 IEEE international energy efficiency in buildings. Pacific Grove, CA; 2004.
Conference on Control Applications, CCA 2010. Yokohama, Japan: Institute [140] Lee K-H, Braun JE. Development and application of an inverse building model
of Electrical and Electronics Engineers Inc.; 2010. p. 2190–5. for demand response in small commercial buildings. In: Proceedings of
[116] Wang S, Xu X. Simplified building model for transient thermal performance SimBuild; 2004. p. 1–12.
estimation using GA-based parameter identification. Int J Therm Sci [141] Morgan S, Moncef K. Field testing of optimal controls of passive and active
2006;45:419–32. thermal storage. In: Proceedings of the 2010 ASHRAE winter conference.
[117] Pakanen J, Karjalainen S. A state machine approach in modelling the heating Orlando, FL, United states: Amer. Soc. Heating, Ref. Air-Conditoning Eng. Inc.;
process of a building. Energy Build 2009;41:528–33. 2010. p. 134–46.
 D. Lazos et al. / Renewable and Sustainable Energy Reviews 39 (2014) 587–603 603

[142] Greensfeldera EM, Henzea GP, Felsmannb C. An investigation of optimal Congress and Exposition, IMECE 2010. PARTS A AND B. Vancouver, BC,
control of passive building thermal storage with real time pricing. J Build Canada: American Society of Mechanical Engineers; 2010. p. 759–67.
Perform Simul 2011;4:91–104. [154] Yun K, Cho H, Luck R, Mago PJ. Real-time combined heat and power
[143] Wang F, Pichatwatana K, Roaf S, Zhao L, Zhu Z, Li J. Developing a weather operational strategy using a hierarchical optimization algorithm. 4th ed. 55
responsive internal shading system for atrium spaces of a commercial City Road, London, EC1Y 1SP, United Kingdom: SAGE Publications Ltd.; 2011.
building in tropical climates. Build Environ 2014;71:259–74. p. 403–12.
[144] Hu J, Karava P. Model predictive control strategies for buildings with mixed- [155] Zong Y, Mihet-Popa L, Kullmann D, Thavlov A, Gehrke O, Bindner HW. Model
mode cooling. Build Environ 2014;71:233–44. predictive controller for active demand side management with PV self-
[145] Ferguson AB, Winn CB. Development of an optimal controller of a small consumption in an intelligent building. In: Proceedings of the 2012 3rd IEEE
commercial passive solar building. In: Solar engineering 1989 – proceedings PES Innovative Smart Grid Technologies Europe, ISGT Europe 2012. Berlin,
of the eleventh annual ASME solar energy conference. San Diego, CA, USA: Germany: IEEE Power and Energy Society; Technische Universitat Berlin –
American Soc of Mechanical Engineers (ASME); 1989. p. 65–9. SENSE Lab; 2012.
[146] Ferreira PM, Silva SM, Ruano AE. Model based predictive control of HVAC [156] Ferhatbegovic T, Palensky P, Fontanella G, Basciotti D. Modelling and design
systems for human thermal comfort and energy consumption minimisation. of a linear predictive controller for a solar powered HVAC system. In:
In: Proceedings of the 1st IFAC Conference on Embedded Systems, Computa-
Proceedings of the 2012 IEEE 21st International Symposium on Industrial
tional Intelligence and Telematics in Control, CESCIT 2012. Wurzburg,
Electronics, 28-31 May 2012. Piscataway, NJ, USA: IEEE; 2012. p. 869–74.
Germany: IFAC Secretariat; 2012. p. 236–41.
[157] Wahab HA, Duke M, Carson JK, Anderson T. Studies of control strategies for
[147] Hagras H, Packham I, Vanderstockt Y, McNulty N, Vadher A, Doctor F.
building integrated solar energy system. In: Proceedings of the 2011 IEEE 1st
An intelligent agent based approach for energy management in commercial
Conference on Clean Energy and Technology, CET 2011. Kuala Lumpur,
buildings. In: 2008 IEEE International Conference on Fuzzy Systems, FUZZ
Malaysia: IEEE Computer Society; 2011. p. 342–7.
2008. Hong Kong, China: Institute of Electrical and Electronics Engineers Inc.;
[158] Liu S, Henze GP. Evaluation of reinforcement learning for optimal control of
2008. p. 156–62.
building active and passive thermal storage inventory. J Sol Energy Eng, Trans
[148] Kummert M, Andre P, Nicolas J. Optimal heating control in a passive solar
commercial building. In: Proceedings of the EuroSun 2000 ISES-Europe solar ASME 2007;129:215–25.
congress. UK: Elsevier; 2000. p. 103–16. [159] Stadler M, Kloess M, Groissbock M, Cardoso G, Sharma R, Bozchalui MC, et al.
[149] Vakiloroaya V, Samali B, Madadnia J, Ha QP. Component-wise optimization Electric storage in California's commercial buildings. Appl Energy
for a commercial central cooling plant. In: Proceedings of the 37th Annual 2013;104:711–22.
Conference of the IEEE Industrial Electronics Society, IECON 2011. Melbourne, [160] Henze GP, Felsmann C, Knabe G. Evaluation of optimal control for active and
VIC, Australia: IEEE Computer Society; 2011. p. 2769–74. passive building thermal storage. Int J Therm Sci 2004;43:173–83.
[150] Sun Y, Wang S, Huang G. Model-based optimal start control strategy for [161] Candanedo JA, Dehkordi VR, Stylianou M. Model-based predictive control of
multi-chiller plants in commercial buildings. Build Serv Eng Res Technol an ice storage device in a building cooling system. Appl Energy 2013;111:
2010;31:113–29. 1032–45.
[151] Asano H, Imamura E, Sogai S. Optimal planning of cogeneration systems [162] Kawashima M, Dorgan CE, Mitchell JW. Optimizing system control with load
under time-of-use rates. Electr Eng Jpn (Denki Gakkai Ronbunshi) 1994;114: prediction by neural networks for an ice-storage system. ASHRAE Trans
102–10. 1996;102:1169–78.
[152] Cho H, Luck R, Eksioglu SD, Chamra LM. Cost-optimized real-time operation [163] Henze GP, Dodier RH, Krarti M. Development of a predictive optimal
of CHP systems. Energy Build 2009;41:445–51. controller for thermal energy storage systems. In: Proceedings of the 1998
[153] Okamoto S. A case study of energy saving by CCHP technology in a hospital. ASHRAE winter meeting, Part 1 (of 2), Pt 1A. San Francisco, CA, USA:
In: Proceedings of the ASME 2010 International Mechanical Engineering ASHRAE; 1998. p. 54.