Professional Documents
Culture Documents
درجة ثاننية
درجة ثاننية
درجة ثاننية
Applied Energy
journal homepage: www.elsevier.com/locate/apenergy
H I GH L IG H T S
• Developing building cooling forecasting model for energy saving of HVAC system.
• Forecasting step-ahead building cooling load by time series prediction.
• Forecasting step-ahead building cooling load by Fourier’s law-based analysis.
• Improving performance of building cooling load forecasting by ensemble technique.
A R T I C LE I N FO A B S T R A C T
Keywords: Short-term load prediction, which forecasts a building’s thermal load with a lead time ranging from seconds to a
Dynamic load forecasting few days, is essential for not only monitoring and controlling the system operation, but also on-line scheduling.
Building cooling loads Dynamic cooling load forecasting, which belongs to short-term load prediction, is both meaningful for mon-
Artificial neural network itoring the system or fuzzy on-line scheduling and crucial for solving the time-lag problem to meet the heating,
Ensemble approach
ventilation and air-conditioning system’s time-varying cooling loads. Numerous studies have been carried out to
develop dynamic load-forecasting models, and great achievements have been made. However, limitations in
their applicability persist because most previous models are calendar- and time-based data-driven models that
may fail when unexpected issues occur or special schedules are adopted. What’s more, the inputs that were
selected passively from the source data pools at hand rather than via active exploration may be insufficient and
impair the accuracy of forecasting models. This paper proposes a novel dynamic forecasting model for building
cooling loads that combines an artificial neural network with an ensemble approach. Based on physical prin-
ciples other than the available data source, the inputs are explored actively and are independent from both
calendar and time indicators, which make the forecasting model being capable of dealing with irregular occa-
sions and unexpected schedules with high accuracy. A benchmark is proposed that uses the current load Q (t ) as a
forecasted cooling load Q (t + i ) and gives the minimum accuracy requirement for a dynamic forecasting model.
The benchmark not only can be used to evaluate dynamic forecasting models that are validated by various case
studies, but also ensures that the proposed forecasting model can be applied immediately to heating, ventilation
and air-conditioning systems to tackle the time-lag problem.
⁎
Corresponding author.
E-mail address: lawang9-c@my.cityu.edu.hk (L. Wang).
https://doi.org/10.1016/j.apenergy.2018.07.085
Received 3 May 2018; Received in revised form 2 July 2018; Accepted 14 July 2018
0306-2619/ © 2018 Elsevier Ltd. All rights reserved.
L. Wang et al. Applied Energy 228 (2018) 1740–1753
Fig. 1. Diagram of model-based optimisation control. (All figures of this paper are produced by the authors.)
is hardly true in reality since time delay due to the slow response of the unexpected incidents occur. In addition, the inputs selected from the
air-side of the HVAC system cannot be avoided. The step-wise opti- source data pools at hand may be insufficient. The use of inputs filtered
mised setting based on the current load would make the HVAC system from the source data pools would be reasonable only if the source data
fail to achieve maximum coefficient of performance (COP) in the pools included all necessary factors. If the source data pools failed to
coming timestep. This is the so-called time-lag problem as demon- include necessary inputs, the filtered inputs would be even more in-
strated in Fig. 2, which also impairs the energy efficiency of an HVAC sufficient. The insufficiency of inputs may harm the forecasting model’s
system. For HVAC systems that work for years, tremendous energy performance. In addition, no uniform benchmark can compare all kinds
would be saved if this time-lag problem can be solved. To solve this of forecasting models that are validated by different case studies.
problem, a dynamic cooling load forecast for the next time interval is Therefore, the potential remains to improve the performance of fore-
necessary. casting models, and a reasonable benchmark should be proposed.
Numerous studies have examined this topic, and great achievements This paper proposes a novel dynamic forecasting model for building
have been made, but limitations to their applicability persist because cooling loads that combines an artificial neural network (ANN) with an
most studies are calendar- and time-based forecasting, which limits the ensemble approach. Based on physical principles other than the avail-
forecasting model to a building schedule and makes it unsuitable when able data source, the inputs are explored actively to be trained by the
Fig. 2. Illustration of time-lag problem in meeting an HVAC system’s time-varying cooling load.
1741
L. Wang et al. Applied Energy 228 (2018) 1740–1753
proposed data-driven model. The inputs, which are independent from [16]. A general regression neural network was applied for hourly load
either calendar or time, make the forecasting model capable of dealing forecasting 1 day in advance [17]. An ANN and a fuzzy expert system
with unexpected schedules with higher accuracy. The benchmark, were combined to perform hourly load forecasting 1 day in advance
which requires the forecasted cooling load to be more accurate than [18]. A nonlinear autoregressive neural network model was developed
simply taking the current cooling load as the target, is the minimum to forecast the hourly electric load 1 day in advance [19]. Two linear
accuracy requirement for a forecasting model. Exceeding the bench- regression predictors were combined with an ANN using a Bayesian
mark guarantees that the proposed forecasting model can be applied method to forecast the next day’s load at a given hour [20]. An auto-
immediately to HVAC systems to tackle the time-lag problem. regressive-moving average with an exogenous terms model was com-
In Section 2, the previous practices in short-term load forecasting bined with fuzzy logic to predict the hourly load 1 day in advance [21].
are reviewed. The dynamic load-forecasting model proposed in this AI methods have become increasingly popular in short-term load
paper is introduced in Section 3. A case study then validates the pro- forecasting with the boom of a variety of advanced AI techniques.
posed dynamic load-forecasting model in Section 4, and conclusions are Although ANNs remain popular, all kinds of other techniques were
drawn in Section 5. widely used, such as combination of rough set theory with an ANN and
data fusion, clustering algorithm, support vector regression, support
2. Review of short-term load forecasting vector machine and deep learning.
The support vector machine was adopted to predict the mean
To the authors’ knowledge, the study of short-term load prediction monthly load by determining the mean monthly dry bulb temperature,
can be traced to around the 1970s, and in the early literature the load the relative humidity and global solar radiation [22]. An ANN was
mainly refers to a building’s electric load, including not only the HVAC applied to forecast the hourly electric demands of a chiller with a lead
system but also lighting and equipment. It is difficult to rely on a time of one time step and 1 day, respectively [23–25]. An ANN was
physics-based model to forecast the load because it is usually unrealistic combined with a data-fusion technique to predict the hourly load [26].
to obtain sufficient necessary detailed information. Therefore, several A clustering-based method was combined with a multi-layer perceptron
kinds of data-driven methods that require much less information than to perform short-term load forecasting [27]. The support vector re-
physics-based methods have been discussed in studies. gression was applied to predict the hourly electrical load of residential
The previous studies can be reviewed from two aspects: the tech- buildings months in advance [28]. Two types of support vector re-
niques used to deal with data and the parameters selected as the inputs gression technique were united to forecast the half-hourly and daily
for the data-driven models. These two aspects are reviewed in Sections energy consumption of buildings [29]. A support vector machine and
2.1 and 2.2, respectively. the deep learning technique were adopted to predict cooling load
profiles 24 h in advance [30]. The enriched AI technique spawned
2.1. Evolution of modelling techniques various forecasting models, to select appropriate models based on
building characteristics, Cui et al. [31] proposed a meta-learning based
Modelling techniques have changed from statistical methods to ar- framework that effectively recommended most appropriate load fore-
tificial intelligence (AI) methods. Before the emergence of the AI casting model for each unique building.
method, around 1966 to 1990, statistical approaches were the most In summary, the evolution of modelling techniques used for short-
popular techniques in the short-term load-forecasting model [1,4,5]. term load forecasting has progressed from statistical techniques to AI
The AI method’s application in short-term load-forecasting models techniques. The AI techniques, which adapt themselves by learning
began with its development in the 1990 s. The transition from statistical from the data samples, release the forecasting model from elaborate
methods to AI methods occurred around 1991 to 2001, and with the data pre-analysis. With the application of AI techniques, the load profile
booming of advanced AI techniques, a variety of AI methods and hy- can be disposed as a whole rather than being disassembled as different
brids of various AI methods have been widely used ever since [6–8]. components as before, and the forecasting method is more concise.
Multiple regression methods, exponential smoothing and stochastic
time series are popular statistical approaches that have been applied in 2.2. Selection of model inputs
short-term load forecasting. Multiple linear regression was used to
forecast the daily peak load (i.e. the hourly integrated load at 2 PM) Except for the building properties, which usually remain stable over
[9,10]. General exponential smoothing was used for hourly load fore- time, the relevant factors that affect a building’s cooling load are the
casting with a lead time of 1–24 h [2]. The long-term average was weather conditions, occupancy and operating equipment in the
combined with exponential smoothing to conduct hourly load fore- building. Due to the difficulty in collecting occupancy information, time
casting for the following day [11]. The stochastic models were adopted indicators – time of day, day of the week, day type (holiday or not) – are
for hourly power system load forecasting with lead times from 1 to 24 h usually chosen as inputs to represent occupancy scenarios. In addition,
[12], and the stochastic time series analysis was applied to predict the the historical load is a popular input because it indicates the trend of
daily peak load 1 day ahead [13]. the load profile in a mathematical way. Consequently, three types of
The statistical approaches, which call for an explicit mathematical input – weather conditions, time indicators and historical loads – have
model to give the relationship between the output and inputs, were been widely chosen in existing studies to include the effects of the re-
applied with elaborate analysis of the source data. As such, studies that levant factors. Many combinations of these three types of inputs have
use statistical approaches have tended to divide the total load into been widely discussed.
various components by their nature. For example, the total load was For example, only weather data – the external hourly temperature of
divided into a basic load and a weather-sensitive load [9–11]. Besides, a 1 day before –was adopted as input to forecast the hourly load 1 day
weekly cycle component was appended sometime [12]. ahead [17]. Only historical loads were taken as input to forecast hourly
During the transition period from statistical methods to AI methods, loads with a lead time from 1 to 24 h [13,25,32]. Historical loads and
ANNs, fuzzy neural networks, combined ANNs with regression, com- weather data (ambient dry bulb temperature) were combined as input
bined ANNs with fuzzy expert system and other techniques were to forecast the daily peak load, daily total load and hourly load [14].
adopted. A three-layer ANN with historical loads and historical and Historical loads and the type of day were united to forecast the hourly
forecasted dry bulb temperatures was used to predict the daily peak load with a lead time from 1 to 24 h [2,16]. Weather data and the type
load, total load and hourly load [14]. A fuzzy neural approach was of day were taken as input to forecast the daily peak load and hourly
adopted to predict the following day’s load curve [15], and a fuzzy loads, respectively, 24 h in advance [9,33]. The historical hourly load,
neural network was applied to forecast the hourly load curve using weather data and type of day were combined as inputs in several
1742
L. Wang et al. Applied Energy 228 (2018) 1740–1753
1743
L. Wang et al. Applied Energy 228 (2018) 1740–1753
reasonably sufficient. The Fourier’s law-based analysis provides a per- temperature, the humidity, the number of occupants and the outside
spective for tracking the effective factors of building cooling loads. surface temperature of the walls should be considered as inputs to
As shown in Fig. 3, taking a building block with concrete envelopes forecast a building’s cooling load. The building properties that remain
with thickness X and area A as an example, the building cooling load the same over a period are excluded.
can be roughly traced back to four portions as follows. Heat gain from the ambient environment when shaded,
T∞−Tin
(1) The heat transferred from the ambient air to the indoor air, as- Qe = R1 R2 R 4 R5
+ R3 + (2)
suming that the building block is shaded from sunlight. The heat R1 + R2 R 4 + R5
flow is first transferred from the ambient air to the outside surface
where R1 is affected by the ambient humidity [39],
of the walls via radiation and convection, from the outside surface
to the inside surface of the walls via conduction, and finally to the R1 = f (H∞) (3)
indoor air via radiation and convection (see Eq. (2)). The radiant
heat resistance of the ambient air is affected by the humidity and Heat gain from the fresh air,
can be described as a function of the humidity (see Eq. (3)). Qs, f = Cp ṁ air (T∞−Tin ) (4)
(2) The heat gain from fresh air. Fresh air is necessary for the occu-
pants’ health and contributes to the cooling load via sensible and Ql, f = ṁ air (H∞−Hin ) Lc (5)
latent heat gain. The sensible heat gain from fresh air can be cal-
culated with Eq. (4), and the latent heat gain can be calculated with ṁ air = f (p) (6)
Eq. (5). The necessary mass flow rate of fresh air is closely related to
where p is the number of occupants.
the number of occupants (see Eq. (6)).
Latent heat gain from occupants,
(3) Latent heat gain from the occupants. The occupants’ main con-
tribution to the cooling load is via sweat evaporation and is roughly Qp = f (p) (7)
a function of the occupants’ number for a building with normally
one type of activity (see Eq. (7)). Heat gain from solar radiation,
(4) The heat gain from solar radiation, which is reflected mainly by the Qsol = f (Toutsurf ) (8)
outside surface temperature of the walls (see Eq. (8)).
where Toutsurf is the outside surface temperature of the walls.
To sum up the factors that affect the building cooling load men- Summary of effect factors of building cooling loads,
tioned above, a function can be written as Eq. (9). The ambient dry bulb
Q = f (building properties, p , T∞, H∞, Toutsurf ) (9)
1744
L. Wang et al. Applied Energy 228 (2018) 1740–1753
3.2. Dynamic load-forecasting model combining artificial neural network current building cooling load and wet bulb temperature as constraints
with ensemble approach [50].
Because the current load is used as the average load of the coming
The theoretical basis illustrated in Section 3.1 provides the selection time interval in real controls, the accuracy of the forecasted load of the
of inputs related to building cooling loads from two perspectives, based coming time interval should be more accurate than that obtained by
on which two forecasting models – Models A and B – can be derived. taking the current load as a forecast. Consequently, the benchmark of
The relationship between the inputs and cooling loads can be estab- forecasting with the lead time of a single time interval i would be the
lished by an ANN, which is a dynamic inverse model that learns from same as taking the current cooling load Q (t ) as a forecast of the next
examples and can approximate the behaviour of highly nonlinear sys- time interval Q (t + i ) . The benchmark is applied in Section 4 to eval-
tems [40]. ANNs have been widely used in short-term load forecasting uate the performance of the proposed dynamic forecasting model.
[8] and have a better architecture to evaluate, estimate and predict a
set of large data than statistical methods [7]. ANN with one hidden 3.3.2. A comparison model
layer is adopted in this study, since the output of an ANN model with Besides the benchmark, a model for performance comparison,
single hidden layer is in fact a superposition of numerous weighted which is a popular dynamic load forecasting models, is built up to
sigmoid functions which has been proven to be a universal function further evaluate the performance of the proposed load forecasting
approximator [41]. model. With 4 previous time step inputs, including cooling loads (KW),
The input variables for Model A are historical cooling loads at time the ambient dry bulb temperature (°C), solar horizontal radiation (W/
t, t−i, t−2i, t−3i, etc., namely Q (t ), Q (t −i), Q (t −2i), Q (t −3i), etc. (KW), m2) and room temperature setpoint(°C), a 3-layer ANN model has been
based on the analysis of Section 3.1.1 (see Fig. 4). The input variables of developed to perform a one time step ahead building cooling load (KW)
Model B are the ambient dry bulb temperature (°C), the relative hu- forecasting [51]. The transfer function is hyperbolic tangent function in
midity (%), the number of occupants and the outside surface tem- the hidden neurons and a linear function is used in the output neuron.
perature of the walls (°C). The dampening effect of the building’s The training function is the Levenberg-Marquardt back- propagation
thermal mass and the accumulation of heat gain from the air and the algorithm.
occupants over time make it necessary to consider the historical value
of each input at time t, t−i, t−2i, t−3i, etc. (see Fig. 5). When time t is
4. Case study
the current time, the output is a forecast of the load for a coming time
interval Q (t + i) (KW).
A building, whose parameters were derived from a real 90-floor
The number of hidden neurons in each simulation is determined by
commercial skyscraper, was developed using the transient system si-
a rule-of-thumb developed by Ward System [42] described as follows.
mulation tool (TRNSYS). The simulation was verified by comparing the
Ninp + Noutp simulated building cooling loads with the measured loads. The simu-
N= + Ns
2 (10) lation overcame the limitations of measured data by providing suffi-
where N is the number of neurons; Ninp and Noutp are the number of cient information such as the envelopes’ outside surface temperature
input and output parameters, respectively; and Ns is the number of and by avoiding unreasonable building cooling loads due to improper
training samples. manipulations of the HVAC system by the occupants. Based on the si-
Models A and B emphasise different aspects of the cooling load mulated data, the proposed dynamic load-forecasting model was suc-
composition and are developed independently based on diverse the- cessfully validated.
ories. The combination of Models A and B by the ensemble approach is
supposed to give more accurate forecasts. The ensemble approach is an 4.1. Building description
idea that by the combination of different predictors, it is possible to
improve the overall prediction accuracy [43]. This idea has been used The sample building of this study is a mega grade A office building,
to improve the performance of prediction models successfully in several which is located at the hub centre of Hong Kong. Multi-national fi-
researches [44–46]. The technique of the ensemble approach used in nancial companies coming from all over the world are working in this
this paper is the bagging approach [47], which averages the inputs to
give outputs.
Concretely, Models A and B are trained by the relevant historical
data, which cover the latest weather and occupancy scenarios. The
building properties are assumed to be in the same state for a period, at
least a few months. After Models A and B are trained, they can be used
for forecasting. With relevant inputs, cooling loads are forecast by
Models A and B, and the final forecasted cooling load can be obtained
using the ensemble approach (see Fig. 6).
3.3.1. Benchmark
The current load combining with the current weather information
are usually taken as constraints to derive the optimum HVAC system
settings to minimize the electrical energy consumption [48].
Current wet bulb temperature and building cooling load are taken as
constraints in optimising the condenser water loop of HVAC system [3].
An energy optimization methodology was proposed to derive the op-
timized operation decisions for chiller plants at regular intervals based
on the building thermal load and the weather condition [49]. A varying
searching bounds optimization control method was proposed by taking Fig. 4. Architecture of the ANN for Model A.
1745
L. Wang et al. Applied Energy 228 (2018) 1740–1753
1746
L. Wang et al. Applied Energy 228 (2018) 1740–1753
Table 1 and B, and the latter one third of the data were used to test the per-
Construction layers and thermal properties of the envelope. formance of the dynamic forecasting model. Previous inputs of nine
Construction U-value (W/ g-Value
time steps and three time steps for Models A and B, respectively, were
m2 K) adopted after a series of trials.
An early-stop validation approach showed no significant improve-
Wall/Roof 200 mm concrete 0.580 – ment in the validation error over a predefined number of epochs after
50 mm insulation layer
Window 6 mm glass and 16 mm air and 6 mm 1.26 0.397
the minimum level was reached. The intermediate state of the model
glass with the minimum validation error was selected as the trained model.
One thousand epochs were adopted in this case study (refer to Fig. 12).
As shown in Fig. 13a, the building loads forecasted by Models A and
load of the sample building in the case study. B were distributed around the real load profile with diverse patterns.
During the peak load period, the results of Model A (dashed line with
4.3. Results and discussion hollow squares) tend to be higher than the real load profile (thick solid
line with crosses). Sometimes, the deviations are huge: around 7110,
4.3.1. Results 7146, 7182, 7206 and 7266 h. However, the results of Model B (thin
Overall, 742 valid data samples were derived from the TRNSYS si- solid line with solid triangles) tend to be randomly distributed around
mulation from 6553 to 7294 h. The first two thirds of the data were the real load during the peak load periods. For example, the results of
trained by an ANN with Bayesian regularisation to develop Models A Model B match the real load profile well around 7116, 7137, 7161,
1747
L. Wang et al. Applied Energy 228 (2018) 1740–1753
7233, 7157 and 7263 h; underestimates occur around 7140 and 7143 h,
whilst overestimates occur around 7185, 7191, 7212 and 7236 h.
During the trough load period, the results of Model A and Model B
match the real load profile well.
During the period that loads are increasing or decreasing, the results
of Model A match the real load profile well and Model B performs well
when the load increases whilst underestimates happen when load de-
creases, such as around 7146 and 7170 h.
The differences in the forecasting patterns derived from Models A
and B make the combination of the results of the two models mean-
ingful. The final results (thin solid line with solid circle) were drawn
after using an ensemble approach to average the results of Models A
and B. As shown in Fig. 13b, both Model A’s large deviations during the
peak load periods and Model B’s underestimates during the decreased
Fig. 10. Scatter plot of simulated cooling loads with measured cooling loads. load periods were alleviated after ensemble. As such, the final results
match the real load profile better than the results of either model alone.
The benchmark (dashed line with half hollow square), which takes the
current load as a forecast, plots a profile that behaves as a shift of the
real load profile. Large deviations could not be avoided by the bench-
mark profile when loads varied dramatically as increasing or de-
creasing. As indicated by the value of R2, the proposed forecasting
model considerably improved the accuracy, as compared with the
benchmark.
The load profile forecasted by the comparison model tend to over-
estimate at both peak and trough load periods, see Fig. 13c. The de-
viation can be huge at some hours, such as around 7113, 7167 and
7239 h. With an R2 of 0.8633, the performance of the comparison
model is inferior to that of the proposed forecasting model.
The errors of the forecasting models were analysed by both error
distribution analysis and scatter plot as shown in Fig. 14. From the error
Fig. 11. Error distribution of simulated cooling loads. distribution histograms, it can be seen that the 95% confidence interval
of the error distributions was located in a similar range for Models A
and B. The mean error of the model after ensemble was the average of
the means of Models A and B, yet the deviation (sigma) was smaller
than that of either Model A or B. The 95% confidence interval of the
error distribution of the model after ensemble was only two thirds that
of Models A and B. The mean error of the benchmark was between that
of Models A and B, whilst the deviation was the greatest, and the
benchmark’s 95% confidence interval of error distribution was nearly
double that of the model after ensemble. Both the mean error and the
deviation of the comparison model are larger than those of the model
after ensemble. The 95% confidence interval of error distribution of the
comparison model is around 1.75 times that of model after ensemble.
The scatter plots in Fig. 14 show that, relatively, the deviations are
relatively large at high load period whilst small at low load period for
Model A. For Model B, the deviations occur at both high load period
Fig. 12. Early-stop validation for ANN training. and low load period. After the ensemble procedure, the deviations are
reduced obviously for all load scenarios. The deviations are averagely
distributed at high load period and low load period besides some
1748
L. Wang et al. Applied Energy 228 (2018) 1740–1753
outliers for the benchmark, and the results of the comparison model better than Model A on the inflection points in general, with smaller
tend to overestimate a lot. deviations. However, during the trough loading periods, when the en-
vironmental factors remained more stable, Model B performed worse
than Model A with a few underestimates.
4.3.2. Discussion The diversity of Models A and B made it reasonable to combine
As shown in Fig. 13a and summarized in Table 2, the building loads them to form a better forecasting model, which was proved by the re-
forecasted by Models A and B were distributed around the real load sults of the case study. The load profile showed that most of the large
profile with diverse patterns because they were based on different deviations were eased by the model after ensemble. The R2 values of
theories and stressed different aspects of the cooling load composition. Models A and B were both around 0.93, indicating similar accuracies of
Based on the RTS method, Model A stressed the radiant heat gain the two models. The R2 value increased to around 0.96 by combining
portion, which depended completely upon the cooling loads that al- the two models, and such a large improvement in accuracy could not
ready existed from the current time back into the past whilst the future occur if the two models were not complementary. The error deviations
was excluded. With the adoption of Model A, the cumulative effects of of Models A and B were both around 830 KW, whilst the residual de-
the cooling loads were well presented, but the inflection points that viation of the model after ensemble was reduced to around 630 KW.
were driven mainly by sudden changes in the weather or occupancy The decrease in the residual deviation also indicated the com-
were hardly reflected. Therefore, huge deviations mainly occurred with plementary status of Models A and B. Compared with the benchmark as
the adoption of Model A for forecasting when the real load profile came well as the comparison model, the proposed ensemble model greatly
to an inflection point on each day, changing from ascending to des- improved the forecasting accuracy. The residual range of the model
cending. During the trough load periods, when the real load profile after ensemble was also considerably smaller than that of the bench-
changed much more moderately, Model A performed better. mark and the comparison model.
Based on Fourier’s law, Model B stressed both the conductive and The forecasting model proposed in this paper differed from previous
convection heat gain portions by considering all relevant environ- studies in input selection. As stated in the Introduction section, inputs
mental factors, indicating the heat gain that affected instant and future in previous studies were usually selected from available source data
cooling loads. The current and historic sensor-based inputs affected pools. A passive method that contained the implicit premise that the
future cooling loads due to the time they consumed with either the available data pools included all necessary inputs, if not, the inputs
dampening effect of the building thermal mass or the process of me- selected from the data pools would have been even more insufficient.
tabolism. Concretely, the outside surface temperature of the envelope However, the accessible source of data pools might not be sufficient for
affected the future cooling load due to the time consumed by con- cooling load prediction, as they were usually collected by the bureau
ductive heat transfer through the envelope’s thermal mass; the ambient according to the basic requirements of weather sampling. Instead of
dry bulb temperature and the relative humidity influenced the future passive selection of relevant inputs based on the accessible data pool,
cooling load as fresh air with the time needed to disperse evenly into a extra inputs were actively explored based on physical modelling.
large room. The number of occupants affected the future cooling load Without a presumed condition, the inputs explored actively based on
with the time required by human metabolism to contribute latent heat physical modelling are possibly be used to supplement the source data.
gain. For the essence of stressing instant and future heat gains by Occupancy, rather than the type of day, was chosen to be one input,
considering all environmental affecting factors, Model B performed
1749
L. Wang et al. Applied Energy 228 (2018) 1740–1753
which enriched the forecasting model to handle unexpected schedules. of occupants could be monitored in a variety of ways, such as lift
The use of the type of day to indicate the occupancy scenario had electricity consumption [53], Wi-Fi connections [62] and passive in-
limitations when unexpected incidents occurred and invalidated the frared sensors [28].
usual schedule. With the development of sensor techniques, the number In addition, the envelope’s outside surface temperature was used as
(a1) (a2)
(b1) (b2)
(c1) (c2)
Fig. 14. Forecasting error distributions.
1750
L. Wang et al. Applied Energy 228 (2018) 1740–1753
(d1) (d2)
(e1) (e2)
Fig. 14. (continued)
Table 2
Summary of forecasting performance of Model A and Model B.
Forecasting performance of Model A Forecasting performance of Model B
Inflection points From ascending to stable or descending Mainly overestimates Mainly match with a few overestimates
From descending to stable or ascending Mainly match Mainly match
an input for the first time. The surface temperature, which reflects the 5. Conclusions
heat gain from solar radiation and the heat loss due to wind speed, has
been neglected in other studies. Compared with the use of solar ra- A dynamic load-forecasting model that combines two sub-models –
diation and wind speeds as inputs, the surface temperature is more Models A and B – has been proposed in this paper using an artificial
closely related to the building’s thermal behaviour and can ease com- neural network and an ensemble approach. Based on the radiant time
putation loads by reducing the number of inputs and can be collected series method, Model A stresses the historical radiant heat gain portion
with sensors. of cooling loads; and according to Fourier’s law, Model B emphasises
The framework of the proposed model differed from those in pre- the instantaneous and future conduction and convection heat gain
vious studies. Although ANNs are frequently used in dynamic load portions of cooling loads. By bringing together the complimentary sub-
forecasting, the ensemble approach was adopted here for the first time models, all the heat gain portions that contribute to building cooling
to combine two complimentary models. loads are covered in the dynamic load-forecasting model, and it pro-
The forecasting model proposed in this paper essentially used an vides improved accuracy compared with previous popular forecasting
ANN and the ensemble approach to determine how a building trans- model (the comparison model). It was also demonstrated by the
forms different portions of heat gain – radiation, conduction and con- benchmarking test which represents an ordinary operation in practice.
vection – to the building cooling load with assumption that the building The proposed Model A and Model B are based on structured analysis
thermal properties do not change much with time. By applying the of the widely used theories for cooling load calculation rather than
framework of the proposed model, the qualified training data are ne- relying heavily on data manipulation. Based on radiant time series
cessary. According to the previous study that evaluated a variety of AI method, Model A generally provides highly accurate forecasting whilst
methods [63], the more qualified training data are accessible, the better overestimates at inflection points; based on Fourier’s Law, Model B
the model performance will be. achieves high accuracy generally with underestimates at trough load
The proposed forecasting model achieves accurate and reliable one period. The forecasting accuracy after the ensemble of Model A and
step-ahead forecasting for building cooling load, which can be applied Model B further improve the forecasting performance than each in-
to tackle the time-lag problem of HVAC system control, the HVAC dividual model with an obvious alleviation of overestimates and un-
system optimisation as well as other building management tasks. derestimates. The theoretical bases–radiant time series method and
Fourier’s law, which have been proved to be effective on calculating
1751
L. Wang et al. Applied Energy 228 (2018) 1740–1753
building cooling loads and heat transfer problems, indicate the feasi- [17] Ben-Nakhi A, Mahmoud M. Cooling load prediction for buildings using general
bility of Model A and Model B on forecasting cooling loads for a variety regression neural networks. Energy Convers Manage 2004;45:2127–41.
[18] Kim K, Park J, Hwang K, Kim S. Implementation of hybrid short-term load fore-
of buildings, so as the model after ensemble. casting system using artificial neural networks and fuzzy expert systems. IEEE Trans
By active exploration of model inputs, the number of occupants and Power Syst 1995;10(3).
outside surface temperature of the envelope are included as inputs. By [19] Chow T, Leung C. Nonlinear autoregressive integrated neural network model for
short-term load forecasting. IEE Proc – Generat Transm Distrib 1996;143(5).
applying the number of occupants rather than a calendar or time-based [20] Kiartzis S, Kehagias A, Bakirtzis A, Petridis V. Short term load forecasting using a
input, the Model B breaks the limitation of specific building schedules Bayesian combination method. Electr Power Energy Syst 1997;19(3):171–7.
and is capable of dealing with irregular occasions and unexpected [21] Yang H, Huang C. A new short-term load forecasting approach using self-organizing
fuzzy ARMAX models. IEEE Trans Power Syst 1998;13(1).
schedules. A benchmark was formulated to evaluate the performance of [22] Dong B, Cao C, Lee S. Applying support vector machines to predict building energy
the proposed forecasting model based on practical operation, and it can consumption in tropical region. Energy Build 2005;37:545–53.
also be used to evaluate other dynamic forecasting models that are [23] Yang J, Rivad H, Zmeureanu R. On-line building energy prediction using adaptive
artificial neural networks. Energy Build 2005;37:1250–9.
validated by different case studies.
[24] Gonzalez P, Zamarreno J. Prediction of hourly energy consumption in buildings
The proposed dynamic load-forecasting model is practically ap- based on a feedback artificial neural network. Energy Build 2005;37:595–601.
plicable since it can be applied to tackle the time-lag problem in HVAC [25] Wang J, Wang J, Li Y, Zhu S, Zhao J. Techniques of applying wavelet de-noising
system control. The time-lag problem, which means the cooling load into a combined model for short-term load forecasting. Electr Power Energy Syst
2014;62:816–24.
cannot be reflected by the temperature change instantly due to the time [26] Hou Z, Lian Z, Yao Y, Yuan X. Cooling-load prediction by the combination of rough
delay in air mixing, would make the cooling load be either over-met or set theory and an artificial neural-network based on data-fusion technique. Appl
un-met by the HVAC system. The problem can be alleviated by in- Energy 2006;83:1033–46.
[27] Tang F, Kusiak A, Wei X. Modeling and short-term prediction of HVAC system with
tegrating the proposed dynamic load forecasting model into the HVAC a clustering algorithm. Energy Build 2014;82:310–21.
system control. For HVAC systems that work for years, large amount of [28] Massana J, Pous C, Burgas L, Melendez J, Colomer J. Short-term load forecasting for
energy could possibly be saved. It can also be applied to other building non-residential buildings contrasting artificial occupancy attributes. Energy Build
2016;130:519–31.
management operations such as optimization control as well as fault [29] Zhang F, Deb C, Lee S, Yang J, Shah K. Time series forecasting for building energy
detection. With a reasonable one time step-ahead load forecasting, the consumption using weighted support vector regression with different evolution
system setting for chiller plants’ optimization can be developed in ad- optimization technique. Energy Build 2016;126:94–103.
[30] Fan C, Xiao F, Zhao Y. A short-term building cooling load prediction method using
vance to minimize the electrical energy consumption. The deviation deep learning algorithms. Appl Energy 2017;195:222–33.
between the forecasted loads with the observed loads can also be used [31] Cui C, Wu T, Hu M, Weir JD, Li X. Short-term building energy model re-
as an indicator to indicate the anomalies in building operation. commendation system: a meta-learning approach. Appl Energy 2016;172:251–63.
[32] Xu X, Huang G, Liu H, Chen L, Liu Q. The study of the dynamic load forecasting
Developing more dynamic load forecasting models to further enhance
model about air conditioning system based on the terminal user load. Energy Build
the performance by ensemble technique and integrating the proposed 2015;94:263–8.
dynamic load-forecasting model into HVAC system control to save en- [33] Powell K, Sriprasad A, Cole W, Edgar T. Heating, cooling and electrical load fore-
ergy would be the future works of this study. casting for a large-scale district energy system. Energy 2014;74:877–85.
[34] Sun Y, Wang S, Xiao F. Development and validation of a simplified online cooling
load prediction strategy for a super high-rise building in Hong Kong. Energy
Acknowledgement Convers Manage 2013;68:20–7.
[35] Wong S, Wan K, lam T. Artificial neural networks for energy analysis of office
buildings with daylighting. Appl Energy 2010;87:551–7.
This work described in this paper was fully supported by a grant [36] Lim H, Kim G. Prediction model of cooling load considering time-lag fore pre-
from the Research Grant Council of the Hong Kong Special emptive action in buildings. Energy Build 2017;151:53–65.
Administrative Region, China [Project No. CityU 11257816]. [37] Pedersen C, Fisher D, Liesen R. Development of a heat balance procedure for cal-
culating cooling loads. ASHRAE Trans 1997;103(Part 2):459–68.
[38] Spitler J, Fisher D, Pedersen C. The radiant time series cooling load calculation
References procedure. ASHRAE Trans 1997;103(Part 2):503–15.
[39] American Society of Heating, Refrigerating and Air-Conditioning Engineers..
ASHRAE handbook: fundamentals. Inch-pound ed. Atlanta (GA): ASHRAE; 2013. p.
[1] Gross G, Galiana F. Short-term load forecasting. Proc IEEE 1987;75(12):1558–73.
4.15–6.
[2] Christiaanse W. Short-term load forecasting using general exponential smoothing.
[40] Zhang G, Patuwo B, Hum Y. Forecasting with artificial neural networks: the state of
IEEE Trans Power Apparat Syst 1971;PAS-91(5).
the art. Int J Forecast 1998;14:35–62.
[3] Thangavelu S, Myat A, Khambadkone A. Energy optimization methodology of multi-
[41] Cybenko G. Approximation by superpositions of sigmoidal function. Math Control
chiller plant in commercial buildings. Energy 2017;123:64–76.
Signals Syst 1989;2(4):303–14. https://doi.org/10.1007/BF02551274.
[4] Mahmoud A, Comerford R, Adams J, Dawson E. Load forecast bibliography. IEEE
[42] NeuroShell 2 manual. Frederick (MA): Ward Syst. Group Inc.; 1995. Ward.
Trans Power Apparat Syst 1980;PAS-99(1).
[43] Carney J, Cunningham P. Tuning diversity in bagged ensembles. Int J Neural Syst
[5] Mohamed A, Naresh S. Short-term load demand modeling and forecasting: a review.
2000;10(4):267–79.
IEEE Trans Syst Man Cybern 1982;SMC-12(3).
[44] Leung P, Lee E. Estimation of electrical power consumption in subway station de-
[6] Zhao H, Magoules F. A review on the prediction of building energy consumption.
sign by intelligent approach. Appl Energy 2013;101:634–43.
Renew Sustain Energy Rev 2012;16:3586–92.
[45] Yuen J, Lee E, Lo S, Yuen R. An intelligence-based optimization model of passenger
[7] Kumar R, Aggarwal R, Sharma J. Energy analysis of a building using artificial neural
flow in a transportation station. IEEE Trans Intell Transport Syst 2013;14(3).
network: a review. Energy Build 2013;65:352–8.
[46] Yuen J, Lee E, Lam W. An intelligence-based route choice model for pedestrian flow
[8] Srivastava A, Pandey A, Singh D. Short-term load forecasting methods: s review.
in a transportation station. Appl Soft Comput 2014;24:31–9.
International conference on emerging trends in electrical, electronics and sustain-
[47] Jain A, Dube R, Chen C. Bootstrap techniques for error estimation. IEEE Trans
able energy systems (ICETEESES-16). 2016.
Pattern Anal Mach Intell 1987;PAMI-9(5):628–33.
[9] Heinemann G, Nordman D. The relationship between summer weather and summer
[48] Chua K, Chou S, Yang W, Yan J. Achieving better energy-efficient air conditioning:
loads – a regression analysis. IEEE Trans Power Apparat Syst 1966;PAS-85(11).
a review of technologies and strategies. Appl Energy 2013;104:8–104.
[10] Corpening S, Reppen N, Ringlee R. Experience with weather sensitive load models
[49] Lu L, Cai W, Soh Y, Xie S, Li S. HVAC system optimization-condenser water loop.
for short and long term forecasting. IEEE Trans Power Apparat Syst 1973;PAS-
Energy Convers Manage 2004;45(4):613–30.
92(6).
[50] Wang L, Lee E, Yuen R. A practical approach to chiller plants’ optimisation. Energy
[11] Lijesen D, Rosing J. Adaptive forecasting of hourly loads based on load measure-
Build 2018;169:332–43.
ments and weather information. IEEE Trans Power Apparat Syst 1971;PAS-90(4).
[51] Li Z, Huang G. Re-evaluation of building cooling load prediction models for use in
[12] Gupta P, Yamada K. Adaptive short-term forecasting of hourly loads using weather
humid subtropical area. Energy Build 2013;62:442–9.
information. IEEE Trans Power Apparat Syst 1972;PAS-91(5).
[52] Buildings Department. Code of practice for fire safety in buildings. Buildings
[13] Goh T, Ong H, Lee Y. A new approach to statistical forecasting of daily peak power
Department; 2011. p. 25–7.
demand. Electr Power Syst Res 1986;10:145–8.
[53] Kwok S, Lee E. A study of the importance of occupancy to building cooling load in
[14] Park D, El-Sharkawi M, Marks R, Atlas L, Damborg M. Electric load forecasting
prediction by intelligent approach. Energy Convers Manage 2011;52:2555–64.
using an artificial neural network. IEEE Trans Power Syst 1991;6(2).
[54] University of Wisconsin-Madison. TRNSYS 16: A TRaNsient SYstem Simulation
[15] Srinivasan D, Liew A, Chang C. Forecasting daily load curves using a hybrid fuzzy-
Program – Volume 5 mathematical reference. Madison: University of Wisconsin;
neural approach. IEEE Proc-Generat Transm Distrib 1994;141(6).
2006. p. 137–40.
[16] Bakirtzis A, Theocharis J, Kiartizis S, Satsios K. Short term load forecasting using
[55] A TRaNsient SYstems Simulation Program. Solar Energy Laboratory University of
fuzzy neural network. IEEE Trans Power Syst 1995;10(3).
Wisconsin; 2003.
1752
L. Wang et al. Applied Energy 228 (2018) 1740–1753
[56] Loutzenhiser P, Manz H, Felsmann C, Strachan P, Frank T, Maxwell G. Empirical Build 2015;86:93–103.
validation of models to compute solar irradiance on inclined surfaces for building [61] Ruiz-Calvo F, Montagud C, Cazorla-Marin A, Corberan J. Development and ex-
energy simulation. Sol Energy 2007;81:254–67. perimental validation of a TRNSYS dynamic tool for design and energy optimization
[57] Al-ajmi F, Hanby V. Simulation of energy consumption for Kuwaiti domestic of ground source heat pump systems. Energies 2017;10:1510. https://doi.org/10.
buildings. Energy Build 2008;40:1101–9. 3390/en10101510.
[58] Chargui R, Sammouda H. Modeling of a residential house coupled with a dual [62] Martani C, Lee D, Robinson P, Britter R, Ratti C. ENERNET: Studying the dynamic
source heat pump using TRNSYS software. Energy Convers Manage relationship between building occupancy and energy consumption. Energy Build
2014;81:384–99. 2012;47:584–91.
[59] Cacabelos A, Eguia P, Migues J, Granada E, Arce M. Calibrated simulation of public [63] Banko M, Brill E. Scaling to very very large corpora for natural language dis-
library HVAC system with a ground-source heat pump and a radiant floor using ambiguation. ACL’01 Proceedings of 39th annual meeting on association for com-
TRNSYS and GeoOpt. Energy Build 2015;108:114–26. putational linguistics, Toulouse, France 2001. p. 26–33. https://doi.org/10.3115/
[60] Djedjig R, Bozonnet E, Belarbi R. Analysis of thermal effects of vegetated envelopes: 1073012.1073017.
Integration of a validated model in a building energy simulation program. Energy
1753