Conference PaperPDF Available

Prediction of HVAC loads at different spatial resolutions and buildings using deep learning models

Authors:

Abstract and Figures

This paper explores the applicability of deep learning-driven models for the prediction of energy consumption in generic commercial buildings. The modeling approach relies on recurrent neural networks (RNNs), while the input consists of physical data streams such as indoor air temperature in different thermal zones and data obtained from the central heating ventilation and air conditioning (HVAC) system. The research steps include the implementation of an existing RNN-based model for energy consumption and further model optimization using training and validation sets. The final model was evaluated using the data from two datasets. Additionally, the evaluation performance was tested in case of the varied spatial and system granularities. The results showed that the optimal model architecture was dataset-agnostic. The results showed that predicting the HVAC energy consumption is more challenging at the higher spatial granularity, when compared to building wise or multi-zone wise modeling. Key innovations Exploring the impact of the spatial granularity on the predictive performance of data-driven HVAC energy consumption models in commercial buildings. Practical implications The practical implications of this work can be summarized as follows: • Applicability of the model to different energy consumption data sets with no tuning and calibration costs. • Adaptation of the model to the day-ahead estimation of energy consumption data for a better load scheduling and an increased use of renewable energy sources.
Content may be subject to copyright.
Prediction of HVAC loads at different spatial resolutions
and buildings using deep learning models
Antonio Liguori1,, Shiying Yang1,, Romana Markovic2, Thi Thu Ha Dam1,
erˆome Frisch1, Andreas Wagner2, Christoph van Treeck1
1E3D - Institute of Energy Efficiency and Sustainable Building,
RWTH Aachen University, Aachen, Germany
2Karlsruhe Institute of Technology, Karlsruhe, Germany
Equal contribution, corresponding author: liguori@e3d.rwth-aachen.de
Abstract
This paper explores the applicability of deep learning-
driven models for the prediction of energy consump-
tion in generic commercial buildings. The modeling
approach relies on recurrent neural networks (RNNs),
while the input consists of physical data streams such
as indoor air temperature in different thermal zones
and data obtained from the central heating ventila-
tion and air conditioning (HVAC) system. The re-
search steps include the implementation of an exist-
ing RNN-based model for energy consumption and
further model optimization using training and vali-
dation sets. The final model was evaluated using the
data from two datasets. Additionally, the evaluation
performance was tested in case of the varied spatial
and system granularities. The results showed that
the optimal model architecture was dataset-agnostic.
The results showed that predicting the HVAC energy
consumption is more challenging at the higher spa-
tial granularity, when compared to building wise or
multi-zone wise modeling.
Key innovations
Exploring the impact of the spatial granularity on the
predictive performance of data-driven HVAC energy
consumption models in commercial buildings.
Practical implications
The practical implications of this work can be sum-
marized as follows:
Applicability of the model to different energy
consumption data sets with no tuning and cal-
ibration costs.
Adaptation of the model to the day-ahead esti-
mation of energy consumption data for a better
load scheduling and an increased use of renew-
able energy sources.
Introduction
The European policies identified energy consumption
in buildings as one of the key goals by 2020 and be-
yond (D’Agostino et al. (2017), P´erez-Lombard et al.
(2008)). In both residential and commercial build-
ings, a significant proportion of energy is consumed
for HVAC systems (P´erez-Lombard et al. (2008)).
At the same time, HVAC systems have the largest
energy-saving potential, ranging between 15-30% for
commercial buildings (P´erez-Lombard et al. (2008),
Li and Wen (2014)).
The buildings’ energy saving potential can be fulfilled
by two mechanisms. The first approach relies on the
improvements in technical equipment and more ef-
ficient building operation. Additionally, the second
approach focuses on a better use of the renewable
energy, while considering the fluctuating nature of
renewable energy sources and challenges related to
energy storage. In summary, both latter mechanisms
for improving energy efficiency require reliable pre-
diction of the building’s energy consumption. Here,
data-driven models can be seen as a promising tool.
This motivated a number of studies on data-
driven modeling of energy consumption (Kusiak
et al. (2010), Maasoumy and Sangiovanni-Vincentelli
(2012), Ahmad et al. (2017), Capozzoli et al. (2017),
Markovic et al. (2020), Koschwitz et al. (2018)). Ku-
siak et al. (2010) used a data-driven approach to
minimize the energy usage of air-conditioning in an
office-type facility, while Maasoumy and Sangiovanni-
Vincentelli (2012) presents a hierarchical control
strategy model to balance comfort and energy con-
sumption with a reduction in HVAC energy consump-
tion. Niu et al. (2018) made use of different data-
driven algorithms to estimate HVAC related elec-
tricity consumption in a university dormitory while
O’Neill and O’Neill (2016) created a data-driven ap-
proach using probabilistic graphical theory to predict
building energy performance.
In order to predict total energy consumption of a
building, Li et al. (2017) combined stacked autoen-
coders (SAEs) with the extreme learning machine
(ELM) to benefit from both approaches. Sendra-
Arranz and Guti´errez (2020) developed several long-
short-term memory (LSTM)-based models to predict
the next day energy consumption of a self sufficient
solar house in Madrid, based on the previous day
of measurements. Ben-Nakhi and Mahmoud (2004)
used general regression neural networks (GRNNs) to
predict hourly cooling load for three office build-
ings, while Azadeh et al. (2008), Gonzalez and Za-
marreno (2005), Kim and Cho (2019) and Nizami
and Al-Garni (1995) predicted electric energy con-
sumption using feedforward neural networks. After
achieving accuracy with multi-layer perceptron with-
out transformed variables for estimating energy de-
mand, Pino-Mej´ıas et al. (2017) concluded that the
neural networks are more suitable for combined heat-
ing and cooling energy consumption modeling, when
compared to statistical regression models. When con-
sidering plug loads Markovic et al. (2020) and Zheng
et al. (2019) developed a LSTM model intending to
predict competitively accurate miscellaneous electric
loads (MELs). The results showed that it outper-
formed all other existing MEL models using identical
data sets obtained from commercial buildings, while
Zheng et al. (2019) confirmed that the LSTM archi-
tectures are as well suitable for MEL modeling in res-
idential settings. Wang et al. (2019) also made use of
an LSTM model to predict plug loads and achieved
among linear regression models and traditional arti-
ficial neural networks an improvement in prediction
accuracy.
Grouping different rooms of a building that share
similar characteristics (e.g. orientation, occupancy)
into different thermal zones is a commonly adopted
approach in building performance simulation and it
was confirmed that the right formulation of thermal
multi-zone models can, indeed, lead to lower com-
putational costs without affecting the accuracy of
the building simulation (van Treeck (2010)). On the
other side, whether the influence of varied time res-
olutions has been already evaluated for some appli-
cations (Markovic et al. (2019); Zheng et al. (2019);
Jain et al. (2014)), an analysis of the impact of spa-
tial granularity, i.e. level of aggregation of data, on
the performance of data-driven models is worth fur-
ther investigation. Since there is no generally ac-
cepted definition of spatial granularity (Pozzani and
Zim´anyi (2010)) or spatial resolution, in this con-
text the terms refer to the monitoring of data in a
much more detailed way, such as for a large num-
ber of zones or from a building aggregated to a floor
aggregated scenario. In this respect there is a re-
quirement to keep the practicality and accuracy in
balance (Chong et al. (2021)), because of the many
levels of spatial granularity. In the scope of an early
study, Kalogirou et al. (1997) applied neural networks
to predict space heating loads for spatial granular-
ity varying from small rooms to large spaces of an
educational building in Cyprus. The results proved
that the performance of the proposed model could be
improved by grouping the input data into two dif-
ferent zones. However, due to the lack of historical
data, the inputs consisted mostly of geometrical pa-
rameters. Jain et al. (2014) proposed a sensor-based
forecasting model based on support vector regression
(SVR). This was applied to a multi-family residen-
tial building in New York City in order to estimate
the energy consumption at varied temporal and spa-
tial resolutions. In particular, the spatial resolution
was investigated at building, floor and unit levels. It
was demonstrated that the best model performance
could be achieved at floor level with hourly temporal
granularity. Finally, Zheng et al. (2019) studied the
accuracy of household electricity consumption predic-
tions at different spatial granularity, using LSTM net-
works. The case study consisted of two households
with electricity consumption data aggregated at the
appliance and building-level. The results proved that
prediction accuracy could be higher if the whole en-
ergy consumption of the household was estimated at
the appliance-level (i.e. at the higher granularity).
The following knowledge was gained in the scope of
related existing studies:
The defined spatial resolution has a direct impact
on the goodness of energy consumption predic-
tion (Zheng et al., 2019; Markovic et al., 2020;
Kalogirou et al., 1997; Jain et al., 2014).
The higher spatial resolution leads to better pre-
diction of MEL consumption in both residen-
tal and commercial settings (Zheng et al., 2019;
Markovic et al., 2020).
In contrast to the MEL energy consumption,
the optimal HVAC energy consumption can be
predicted with higher accuracy when applied
to lower spatial granularities (Jain et al., 2014;
Kalogirou et al., 1997).
RNNs and LSTMs neural networks are a suit-
able architecture for predicting both MEL
and HVAC energy consumption (Zheng et al.,
2019; Markovic et al., 2020; Sendra-Arranz and
Guti´errez, 2020).
The energy use patterns in residential buildings
may differ significantly from the ones observed
in commercial buildings (Clevenger et al., 2014).
This work builds up on the existing knowledge re-
garding the suitable model architecture and explore
the impact of the spatial granularity on the HVAC
energy consumption in commercial buildings.
For that purpose, an existing RNN-based model was
implemented and evaluated using two independent
datasets. The original model was further developed
and evaluated using varied spatial data granularity,
from office wise data streams to large office buildings.
The remaining part of the paper is structured as fol-
lows: first, the methodology and the experimental
setting is provided, including the description of the
datasets, optimization approach and evaluation met-
rics. Finally the results are presented and discussed
before the key findings are summarized.
Method and experimental setting
The next section provides the description of the
methodology, including datasets description, opti-
mization approach and evaluation metrics. The ap-
proach followed in this paper is based on the imple-
mentation and optimization of an existing deep learn-
ing model proposed by Chen et al. (2017). The earlier
method was re-implemented and evaluated for model-
ing HVAC loads at different spatial resolutions, using
two datasets.
Datasets
The first dataset was generated using the software
Energy Plus and it was shared by the authors of
the original study (Chen et al. (2017)). The ana-
lyzed building was a 12-storey large office building
located in Seattle, USA, with a total floor area of
46,000 m2. It consisted of 55 input features, which
included zone-level measurements (e.g. indoor air
temperature, relative humidity), centralized HVAC
control and weather variables, recorded in the year
2004. Simulations were performed based on a time
discretization of 10 minutes.
The second dataset consisted of minute-wise moni-
toring data collected in a 3-floor office building lo-
cated in Aachen, Germany. The building had a to-
tal floor area of 7,500 m2with 73 single and dou-
ble occupied offices (F¨utterer and Constantin (2014);
utterer et al. (2013)). The data were collected dur-
ing the year 2014. The minute-wise collected moni-
toring data were down-sampled to a 10 minutes-based
frequency. For further information about the dataset
in question, the reader is referred to F¨utterer and
Constantin (2014); F¨utterer et al. (2013).
Implementation of existing RNN model
The original method proposed by Chen et al. (2017)
consisted of a RNN-based model used to estimate the
total electricity consumption of the previously de-
scribed building in Seattle. The model was firstly
implemented and validated in scope of a round-robin
study, and eventually the further optimization poten-
tial was explored.
The model was based on RNNs, which are typically
used to process sequential data (Goodfellow et al.
(2016)). The main difference with respect to classic
feed-forward configurations is that model’s parame-
ters can be shared among the input time-series, by
operating with a local memory used to store informa-
tion of the past sequence (Chen et al. (2017); Good-
fellow et al. (2016)).
All hyperparameters were fixed as defined in the orig-
inal study. The model’s architecture consisted of one
RNN layer, three feed-forward layers and an output
layer. A dropout is included in every feed-forward
layer with the value of 0.5. Since no information
about the learning rate were provided, it was set to
0.1.
Analogously to the model development presented by
Chen et al. (2017), the training was performed by ran-
domly sampling 10 out of 12 months from the Seattle
dataset and by using the remaining 2 months for per-
formance evaluation. Since the hyperparameters were
fixed as in the original study, no validation set from
the Seattle data was defined. Additionally, electricity
consumption estimation was performed based on the
past 4 hours of input data, for each timestep.
Further model optimization
The previously described model was at first tested us-
ing the data set collected in Aachen and finally fur-
ther optimized. Analogously to the setting used in
the previous steps, the model was further optimized
using a training and a validation set, while it was
tested using an evaluation set. Here, the validation
set was used to identify the optimal model hypothe-
sis, while the evaluation set was used to obtain the
final performance on previously unseen data1. It was
investigated if the use of the longer input sequence
duration could lead to better performance and ex-
periments were conducted where the input sequence
was extended to six hours. Eventually 11 out of 12
months from the Seattle dataset were used as training
and validation sets, while the remaining one month
was used for evaluation. As loss function, the mean
squared error (MSE) was applied and it was mini-
mized using a stochastic gradient descendent (SGD)
optimizer with momentum.
In the second round of experiments, the model was
further optimized on the Aachen dataset, by conduct-
ing hyperparameter tuning using data from all ther-
mal zones from one building level. An overview of the
used features from the Aachen dataset is provided in
Table 1.
Table 1: Overview of the used features for the model
applied on the Aachen dataset.
Feature Unit of measure
Indoor air temperature °C
Set point temperatures °C
(two thermostats per office)
Indoor CO2concentration ppm
Indoor relative humidity %
“Day of the week” index -
Outdoor air temperature °C
Outdoor relative humidity %
Soil temperature over 5cm °C
Liquid precipitation mm
Average wind speed m/s
Wind direction deg
Atmospheric pressure mbar
Global radiation W/m2
Diffuse radiation W/m2
Thermal load kW
1For the formal definition of training, validation and test
sets, the reader is referred to Goodfellow et al. (2016).
The hyperparameter tuning was conducted on the
validation set and the selected optimal hyperparam-
eters are summarized in Table 2. The chosen compu-
tational environment consisted of Python 3.6.8, Ten-
sorflow 1.14.0 and Keras 2.2.0.
Since the total electricity consumption was not avail-
able in the second building, the target variables were
re-defined as heating and cooling loads. This resulted
in two different model’s formulations, namely winter
(W) and summer (S) mode. For both cases, 11 out
of 12 months were used as training and validation
sets, while the remaining month (May and Novem-
ber for summer and winter mode, respectively) was
used as evaluation. The defined data split for each
target variable from the Aachen dataset is presented
in Figure 1.
Table 2: Hyperparameter search results.
Hyperparameter Value
Batch size 200
RNN hidden units 128
RNN hidden layers 1
Feed-forward hidden units 64
Feed-forward hidden layers 3
Learning rate 0.1
Dropout rate 0
Epochs 160
1 2 3 4 5 6 7 8 9 10 11 12
Months
Training Set Validation Set
Evaluation Set
Winter Mode
Evaluation Set
Summer Mode
Figure 1: Training, validation and evaluation data
split (winter and summer mode) for each spatial gran-
ularity.
In addition to exploring the model’s generalization to
different target functions, such as total energy con-
sumption, heating energy consumption and cooling
energy consumption, the applicability of the model
was tested for different spatial resolutions. In partic-
ular, the model was retrained using data that were
aggregated at varied spatial granularity, as in Figure
1. The explored cases included office-wise, floor-wise
and building-wise configurations.
Evaluation metrics
The performance of the proposed method was as-
sessed by using the root mean squared error (RMSE),
the mean absolute error (MAE), the normalized root
mean squared error (NRMSE) and the coefficient of
determination (R2), as formulated by Willmott and
Matsuura (2005) and Hahn (1973). The RMSE was
chosen as the main accuracy indicator, being the es-
tablished evaluation metric by the related research
(Willmott and Matsuura (2005)) and was defined as
follows:
RMSE =sPn
i=1(Xobs
iXpred
i)2
n.(1)
Nonetheless, the RMSE alone does not provide suf-
ficient information, neither about the total average
error nor about its variability along the predictive
horizon (Willmott and Matsuura (2005)) and it is
therefore used as a first comparison with models used
in other studies. In order to account for the ambigu-
ity of information from the RMSE, it was opted for
a further absolute indicator, namely MAE (Willmott
and Matsuura (2005)), formulated as:
MAE =Pn
i=1 |Xobs
iXpred
i|
n.(2)
A comparison between varied spatial resolution con-
figurations was made by means of normalized metrics
(i.e. NRMSE and R2), defined as:
NRM SE =RMSE
Xobs
max Xobs
min
,(3)
R2= 1 Pn
i=1(Xobs
iXpred
i)2
Pn
i=1(Xobs
iXobs
mean)2,(4)
where Xobs
iis the real output value, Xpred
iis the
predicted output value, Xobs
mean is the average of the
real output values, Xobs
max is the maximum real out-
put value, Xobs
min is the minimum real output value
and nis the total number of output values. Due to
the stochastic initialization of the model’s weights, it
was decided to run the evaluations 10 times and to
weight the earlier metrics accordingly.
Results
Round robin evaluation of existing model
The obtained model’s performance in terms of
NRMSE was in the same range as in the study earlier
presented by Chen et al. (2017). The originally con-
firmed NRMSE of 0.078 could be reproduced with a
small difference and the NRMSE obtain in this round
robin study was 0.05. Additionally, performance was
quantified using further metrics that include MAE,
RMSE and R2and the results are presented in Ta-
ble 3. Finally, the simulated and measured electricity
consumption trends shown in Figure 2 confirmed that
the model could accurately forecast the energy values
along all the test set.
Table 3: Round-robin study performance obtained on
the dataset from the building in Seattle.
MAE RMSE NRMSE R2
[MJ] [MJ] [-] [-]
Tot 37.71 45.55 0.05 0.97
Figure 2: Monthly course of simulated and measured
electricity consumption for the whole Seattle dataset.
Model evaluation using an independent
dataset
Table 4 summarizes the performance of the proposed
method for estimating the HVAC energy consumption
using the data collected in Aachen in the case of sum-
mer (S) and winter (W) mode, while the predicted
and measured energy consumption are presented in
Figure 3. It can be observed that the model was
able to capture the total building’s energy demand for
both heating and cooling, showing no significant per-
formance drop in forecasting the two variables. This
is also confirmed by the high R2values, which indi-
cate a relatively accurate fitting of the real thermal
supply trends.
Table 4: Model performance on the whole Aachen
dataset.
MAE RMSE NRMSE R2
[kW] [kW] [-] [-]
S 4.56 5.79 0.07 0.92
W 2.94 3.87 0.06 0.88
Figure 3: Monthly course of simulated and measured
thermal loads on for the Aachen dataset.
The results showed that the deviation between the
measured and predicted thermal load in terms of
MAE were 4.56 kW and 2.94 kW, respectively for
summer and winter mode. The resulting RMSE were
respectively 5.79 kW and 3.87 kW, which indicates
that the cooling energy consumption can be predicted
with lower relative error, while the absolute accuracy
in both summer and winter mode was in a similar
range.
Spatial load granularity versus model perfor-
mance
Table 5 summarizes the performance of the proposed
method for the floor-wise granularity case, in both
winter and summer mode.
Table 5: Floor-wise model performance on the Aachen
dataset.
Floor MAE RMSE NRMSE R2
ID [kW] [kW] [-] [-]
S
EG 1.17 1.56 0.10 0.85
OG1 2.03 2.69 0.07 0.88
OG2 2.46 3.14 0.07 0.89
W
EG 1.04 1.31 0.09 0.69
OG1 1.37 1.85 0.08 0.81
OG2 1.81 2.22 0.07 0.85
Additionally, the simulated and measured trends
from the test set are represented in Figure 4. Here,
”EG”, ”OG1” and ”OG2” stand respectively for the
first, second and third floor of the Aachen building.
It is shown that the model could represent the HVAC
on a building- and floor-wise level with similar per-
formance. The pattern behind the superposition of
several zones could be identified using an identical
modeling approach as the building-wise energy con-
sumption modeling approach.
Finally, the model’s performance for the room-wise
granularity case is presented in Table 6. Due to the
limited space, only the mean result over 73 offices, as
well as the offices with the highest and lowest pre-
dictive error are presented. Further results will be
released afterwards in a public library. Despite the
relatively low absolute errors (i.e. MAE and RMSE),
the R2value was negative for some rooms. According
to Equation 4, this could occur in case the mean of
the real data resulted in a better fit than the estima-
tion of the model, as confirmed in Figure 5.
Table 6: Room-wise model performance on the
Aachen dataset.
Room MAE RMSE NRMSE R2
ID [kW] [kW] [-] [-]
S
max 0.57 0.80 0.23 0.44
mean 0.21 0.33 0.21 0.45
min 0.05 0.11 0.14 -2.62
W
max 0.72 0.96 0.27 0.16
mean 0.15 0.26 0.19 -0.14
min 0.02 0.04 0.06 -0.53
Figure 4: Floor-wise monthly course of simulated and
measured thermal loads for the Aachen dataset.
Figure 5: Monthly course of simulated and measured
thermal loads for a single Aachen office, in case of
negative R2.
Discussion
The aim of this study was to explore the impact of the
spatial granularity on the accuracy of the data-driven
models for energy consumption prediction. The inde-
pendent round-robin study confirmed that the RNN-
based modeling approach proposed by Chen et al.
(2017) is suitable for predicting the building-wise en-
ergy consumption. The results of the latter study
could be reproduced and the obtained NRMSE was
equal to 0.05, namely for 0.01 lower, when compared
to the performance reported by the original study.
This non-significant deviation in the NMSE could be
attributed to the random initialization and updates
in the used Keras and Tensorflow libraries for the im-
plementation of the RNN models.
After validating the existing model, the further po-
tential for improving the model’s performance was
explored. Here, the optimization potential was ana-
lyzed for the inclusion of the longer input sequence
duration, using alternative optimizer for model train-
ing and further hyperparameter tuning. The results
showed that the accuracy could be slightly improved
by using SGD with momentum, while the inclusion of
longer input sequences did not lead to performance
improvements. The results obtained from the re-
peated hyperparameter tuning showed that the per-
formance on different datasets could be optimized by
fine tuning the hyperparameters.
The optimal neural network architecture and model-
ing approach were identical in case of both datasets.
This implies that these design decisions are dataset-
agnostic. Based on these results, it is worth to further
investigate deep learning models for generic energy
consumption predictions and to investigate which de-
sign decisions lead to dataset-agnostic model. The
obtained results could be used to formulate the ini-
tial hypothesis. Namely, an arbitrary neural network
that consist of a layer with units for sequential mod-
eling in addition to a small number of feedforward
layers (such as three layers in the scope of this study)
should provide sufficient learning capacity for predic-
tive energy consumption models.
In contrast to the shared optimal model design and
hyperparameters between different datasets, the re-
sults showed that the chosen spatial granularity has
a direct impact on the predictive performance. The
predictive performance dropped significantly in case
of office-wise HVAC energy consumption prediction,
when compared to the floor-wise or building-wise spa-
tial granularity.
Eventually, the accuracy of the model for various pre-
dictive horizons was further investigated. The anal-
ysis was carried out for the whole Aachen dataset,
up to 12 hours ahead in hourly steps. Due to
space constraints, the results are accessible through a
public repository (TBA). The modeling performance
dropped with the increase in predictive horizon, with
the R2score dropping to 0.51 for five hours ahead en-
ergy consumption modeling. Here, the use of a longer
input sequence duration, namely from 6 hours to 1
week, could lead to better predictive performance as
already observed in Markovic et al. (2020).
Conclusion
The aim of this study was to explore the performance
of a RNN based model for predicting the energy con-
sumption at different spatial granularities and for a
set of target functions such as total energy consump-
tion, HVAC cooling load, and HVAC heating load.
The novelty and the key findings may be summarized
as follows:
It could be hypothesized that narrow neural net-
works with a one hidden layer with sequential
connections and several hidden feedforward lay-
ers represents a dataset-agnostic model architec-
ture for energy consumption prediction.
The general pattern behind explored target func-
tions such as total energy consumption, HVAC
cooling load, and HVAC heating load could be
successfully modeled using identical algorithms,
by repeating the model’s training on each target
domain dataset.
The HVAC loads could be predicted with an
identical performance on the building spatial
granularity and multi thermal zone level (floor-
wise).
The predictive performance drops significantly
at the single thermal zone (office-level) spatial
granularity, which confirmed earlier findings by
Kalogirou et al. (1997).
Acknowledgments
Part of this work was funded by the Deutsche
Forschungsgemeinschaft (DFG, German Research
Foundation) – TR 892/4-1 – and by the German Fed-
eral Ministry of Economics and Energy (BMWi) as
per resolution of the German Parliament under the
funding code 03EN1002A. We thank the EBC Insti-
tute, E.ON ERC at RWTH Aachen University for
providing the monitoring data. This paper benefited
greatly from discussions with members of IEA EBC
Annex 79.
References
Ahmad, M. W., M. Mourshed, and Y. Rezgui (2017).
Trees vs neurons: Comparison between random
forest and ANN for high-resolution prediction of
building energy consumption. Energy and Build-
ings 147, 77–89.
Azadeh, A., S. Ghaderi, and S. Sohrabkhani (2008).
Annual electricity consumption forecasting by neu-
ral network in high energy consuming indus-
trial sectors. Energy Conversion and manage-
ment 49 (8), 2272–2278.
Ben-Nakhi, A. E. and M. A. Mahmoud (2004). Cool-
ing load prediction for buildings using general re-
gression neural networks. Energy Conversion and
Management 45 (13-14), 2127–2141.
Capozzoli, A., M. S. Piscitelli, A. Gorrino, I. Bal-
larini, and V. Corrado (2017). Data analytics for
occupancy pattern learning to reduce the energy
consumption of hvac systems in office buildings.
Sustainable cities and society 35, 191–208.
Chen, Y., Y. Shi, and B. Zhang (2017). Modeling and
optimization of complex building energy systems
with deep neural networks. In 2017 51st Asilomar
Conference on Signals, Systems, and Computers,
pp. 1368–1373. IEEE.
Chong, A., G. Augenbroe, and D. Yan (2021). Occu-
pancy data at different spatial resolutions: Build-
ing energy performance and model calibration. Ap-
plied Energy 286, 116492.
Clevenger, C. M., J. R. Haymaker, and M. Jalili
(2014). Demonstrating the impact of the occupant
on building performance. Journal of Computing in
Civil Engineering 28 (1), 99–102.
D’Agostino, D., B. Cuniberti, and P. Bertoldi
(2017). Energy consumption and efficiency tech-
nology measures in european non-residential build-
ings. Energy and Buildings 153, 72 – 86.
utterer, J., A. Constantin, M. Schmidt, R. Stre-
blow, D. M¨uller, and E. Kosmatopoulos (2013). A
multifunctional demonstration bench for advanced
control research in buildings—monitoring, control,
and interface system. In IECON 2013 - 39th An-
nual Conference of the IEEE Industrial Electronics
Society, pp. 5696–5701.
utterer, J. P. and A. Constantin (2014). Energy
concept for the E.ON ERC main building. Volume
4,9. E.ON Energy Research Center.
Gonzalez, P. A. and J. M. Zamarreno (2005). Pre-
diction of hourly energy consumption in buildings
based on a feedback artificial neural network. En-
ergy and buildings 37 (6), 595–601.
Goodfellow, I., Y. Bengio, A. Courville, and Y. Ben-
gio (2016). Deep learning, Volume 1. MIT press
Cambridge.
Hahn, G. J. (1973). The coefficient of determination
exposed. Chemtech 3 (10), 609–612.
Jain, R. K., K. M. Smith, P. J. Culligan, and J. E.
Taylor (2014). Forecasting energy consumption
of multi-family residential buildings using support
vector regression: Investigating the impact of tem-
poral and spatial monitoring granularity on perfor-
mance accuracy. Applied Energy 123, 168–178.
Kalogirou, S., C. Neocleous, and C. Schizas (1997).
Building heating load estimation using artificial
neural networks. In Proceedings of the 17th inter-
national conference on Parallel architectures and
compilation techniques, Volume 8, pp. 14.
Kim, J.-Y. and S.-B. Cho (2019). Electric energy
consumption prediction by deep learning with state
explainable autoencoder. Energies 12 (4), 739.
Koschwitz, D., J. Frisch, and C. Van Treeck (2018).
Data-driven heating and cooling load predictions
for non-residential buildings based on support vec-
tor machine regression and narx recurrent neural
network: A comparative study on district scale.
Energy 165, 134–142.
Kusiak, A., M. Li, and F. Tang (2010). Modeling and
optimization of HVAC energy consumption. Ap-
plied Energy 87 (10), 3092–3102.
Li, C., Z. Ding, D. Zhao, J. Yi, and G. Zhang (2017).
Building energy consumption prediction: An ex-
treme deep learning approach. Energies 10 (10),
1525.
Li, X. and J. Wen (2014). Review of building energy
modeling for control and operation. Renewable and
Sustainable Energy Reviews 37, 517–537.
Maasoumy, M. and A. Sangiovanni-Vincentelli
(2012). Total and peak energy consumption mini-
mization of building hvac systems using model pre-
dictive control. IEEE Design & Test of Comput-
ers 29 (4), 26–35.
Markovic, R., E. Azar, M. K. Annaqeeb, J. Frisch,
et al. (2020). Day-ahead prediction of plug-in loads
using a long short-term memory neural network.
Energy and Buildings, 110667.
Markovic, R., E. Grintal, A. Nouri, J. Frisch, and
C. van Treeck (2019, Sep). Right on Time : Ex-
ploring Suitable Time Discretization for Occupant
Behavior Co-Simulation. In Proceedings of Build-
ing Simulation 2019 : 16th Conference of IBPSA
: Rome, Italy, 2-4 September 2019 / edited by V.
Corrado and A. Gasparella, pp. 2099–2106. 16th
IBPSA Conference, Rome (Italy), 2 Sep 2019 - 4
Sep 2019.
Niu, F., Z. O’Neill, and C. O’Neill (2018). Data-
driven based estimation of hvac energy consump-
tion using an improved fourier series decomposition
in buildings. In Building Simulation, Volume 11,
pp. 633–645. Springer.
Nizami, S. J. and A. Z. Al-Garni (1995). Forecasting
electric energy consumption using neural networks.
Energy policy 23 (12), 1097–1104.
O’Neill, Z. and C. O’Neill (2016). Development of a
probabilistic graphical model for predicting build-
ing energy performance. Applied energy 164, 650–
658.
erez-Lombard, L., J. Ortiz, and C. Pout (2008). A
review on buildings energy consumption informa-
tion. Energy and buildings 40 (3), 394–398.
Pino-Mej´ıas, R., A. P´erez-Fargallo, C. Rubio-Bellido,
and J. A. Pulido-Arcas (2017). Comparison of lin-
ear regression and artificial neural networks models
to predict heating and cooling energy demand, en-
ergy consumption and CO2 emissions. Energy 118,
24–36.
Pozzani, G. and E. Zim´anyi (2010). Defining spatio-
temporal granularities for raster data. In British
National Conference on Databases, pp. 96–107.
Springer.
Sendra-Arranz, R. and A. Guti´errez (2020). A long
short-term memory artificial neural network to pre-
dict daily hvac consumption in buildings. Energy
and Buildings 216, 109952.
van Treeck, C. (2010, February). Introduction to
building performance modeling and simulation. Ha-
bilitation Thesis, Technische Universit¨at M¨unchen,
Germany.
Wang, Z., T. Hong, and M. A. Piette (2019). Predict-
ing plug loads with occupant count data through a
deep learning approach. Energy 181, 29–42.
Willmott, C. J. and K. Matsuura (2005). Advan-
tages of the mean absolute error (MAE) over the
root mean square error (RMSE) in assessing aver-
age model performance. Climate research 30 (1),
79–82.
Zheng, Z., H. Chen, and X. Luo (2019). Spatial gran-
ularity analysis on electricity consumption predic-
tion using lstm recurrent neural network. Energy
Procedia 158, 2713–2718.
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
Occupancy is a significant area of interest within the field of building performance simulation. Through Bayesian calibration, the present study investigates the impact of the availability of different spatial resolution of occupancy data on the gap between predicted and measured energy use in buildings. The study also examines the effect of occupancy data on the quality of the constructed prediction intervals (PIs) using the Coverage Width-based Criterion (CWC) metric. CWC evaluates the PIs based on both their coverage (correctness) and width (informativeness). This investigation takes the form of an actual building case study, with nine months of hourly measured building electricity use, WiFi connection counts as a proxy for occupancy, and actual weather data. In general, the building energy model’s accuracy improves with the occupancy and plug-loads schedule derived from WiFi data. Specifically, the Coefficient of Variation Root Mean Square Error (CV[RMSE]) reduced from 37% to 24% with an exponential improvement in the PIs quality compared to the results obtained with ASHRAE 90.1 reference schedules. However, the increase in prediction accuracy shrank to 5% CV(RMSE) and a comparable CWC upon calibrating the base loads of the reference schedules. Increasing the spatial resolution from building aggregated to floor aggregated occupancy data worsened the CV(RMSE) and CWC, suggesting trade-offs between parameter uncertainty and model bias/inadequacy. These results contribute to our understanding of the interactions between model complexity, simulation objectives, and data informativeness, facilitating future discussions on the right level of abstraction when modeling occupancy.
Conference Paper
Full-text available
This work describes how to find an optimal temporal discretization of occupant behavior (OB) models with and without building performance simulation (BPS) feedback. Three window opening models proposed by related research were calibrated using monitoring data sampled in varied time frequencies between one minute and one hour. The impact of the time-steps size on the accuracy of the OB models was analyzed using the minute-wise logged ground truth data. Eventually, the analyzed OB models were coupled with a calibrated Modelica-based BPS using a functional mock-up unit (FMU). The results showed that the co-simulation in hourly time-steps could reliably represent the OB in BPS. Furthermore, the analysis of the OB models using different time resolution showed, that given enough monitoring data samples, the sparse temporal resolution acts as a simple but effective regularization strategy in case of data-driven OB models.
Article
Full-text available
The building sector takes a large proportion of electricity consumption and carbon emission in high-density urban areas. To reduce the carbon emissions and use energy more efficiently in the building sector, home energy management system (HEMS) is proposed and used. In the HEMS, the prediction of electricity consumption in the short-term future is used to support the decision makings in the HEMS. Although there existed a number of studies in the prediction of electricity consumption in buildings, there lacks a spatial analysis in the prediction performance, especially on the appliance or sub-meter level and household level. The authors made an assumption that by the performance of household energy consumption prediction can be significantly improved if the prediction is aggregated from the prediction data at the appliance or sub-meter level. Next, two typical datasets are used to validate the assumption by comparing the prediction performance of aggregating the prediction data at appliance level and the one of making direct prediction at the household level. The models used for the prediction are standard stateful long short-term memory (LSTM) neural networks, which have been proofed to be promising in load prediction by previous studies. The results from the comparison validated the assumption, showing that the prediction performance can be significantly improved if prediction is made at the appliance-level first and then aggregated to get the household-level prediction. Therefore, the authors conclude that prediction at the finer appliance granularity level can significantly improve the performance of household-level electricity prediction.
Article
Full-text available
As energy demand grows globally, the energy management system (EMS) is becoming increasingly important. Energy prediction is an essential component in the first step to create a management plan in EMS. Conventional energy prediction models focus on prediction performance, but in order to build an efficient system, it is necessary to predict energy demand according to various conditions. In this paper, we propose a method to predict energy demand in various situations using a deep learning model based on an autoencoder. This model consists of a projector that defines an appropriate state for a given situation and a predictor that forecasts energy demand from the defined state. The proposed model produces consumption predictions for 15, 30, 45, and 60 minutes with 60-minute demand to date. In the experiments with household electric power consumption data for five years, this model not only has a better performance with a mean squared error of 0.384 than the conventional models, but also improves the capacity to explain the results of prediction by visualizing the state with t-SNE algorithm. Despite unsupervised representation learning, we confirm that the proposed model defines the state well and predicts the energy demand accordingly.
Article
The aim of this work is to develop and validate a miscellaneous electric loads (MEL) predictive model that does not require occupant-wise or building-wise model training nor model adaptation while achieving competitive accuracy. For that purpose, a long-short-term memory (LSTM) model was developed using monitored data from a research building located in Abu Dhabi, United Arab Emirates (UAE). In order to test the generalization capabilities of the proposed method, the model was evaluated using data from two additional buildings, a bank office building located in Frankfurt, Germany, and a university building in Ottawa, Canada. The results showed that the developed LSTM is applicable to the tested buildings without the need for occupant-wise or building-wise calibration, hence, addressing an important gap in the existing literature. In addition, a set of MEL predictive models from the literature, that are based on a Weibull distribution and Gaussian mixture models (GMM) are implemented and evaluated using the three identical data sets. The round-robin evaluation of existing MEL predictive models showed that the proposed LSTM model outperformed them especially when a combination of MEL and occupancy information was used as inputs. Finally, the neural network saturation was identified as the key challenge when developing an LSTM-based model for MEL prediction.
Article
In this paper, the design and implementation process of an artificial neural network based predictor to forecast a day ahead of the power consumption of a building HVAC system is presented. The featured HVAC system is situated at MagicBox, a real self-sufficient solar house with a monitoring system. Day ahead prediction of HVAC power consumption will remarkably enhance the Demand Side Management techniques based on appliance scheduling to reach defined goals. Several multi step prediction models, based on LSTM neural networks, are proposed. In addition, suitable data preprocessing and arrangement techniques are set to adapt the raw dataset. Considering the targeted prediction horizon, the models provide outstanding results in terms of test errors (NRMSE of 0.13) and correlation, between the temporal behavior of the predictions and test time series to be forecasted, of 0.797. Moreover, these results are compared to the simplified one hour ahead prediction that reaches nearly optimal test NRMSE of 0.052 and Pearson correlation coefficient of 0.972. These results provide an encouraging perspective for real-time energy consumption prediction in buildings.
Article
Predictive control has gained increasing attention for its ability to reduce energy consumption and improve occupant comfort in buildings. The plug loads prediction is a key component for the predictive building controls, as plug loads is a major source of internal heat gains in buildings. This study proposed a novel method to apply the Long-Short-Term-Memory (LSTM) Network, a special form of Recurrent Neural Network, to predict plug loads. The occupant count and the time have been confirmed to drive the plug load profile and thus selected as the features for the plug load prediction. The LSTM network was trained and tested with ground truth occupant count data collected from a real office building in Berkeley, California. Results from the LSTM network markedly improve the prediction accuracy compared with traditional linear regression methods and the classical Artificial Neural Network. 95% of 1-hour predictions from LSTM network are within ±1 kW of the actual plug loads, given the average plug loads during the office hour is 8.6 kW. The CV(RMSE) of the predicted plug load is 11% for the next hour, and 20% for the next eight hours. Lastly, we compared four prediction approaches with the office building we monitored: LSTM vs. ARIMA, with occupant counts vs. without occupant counts. It was found, the prediction error of the LSTM approach is around 4% less than the ARIMA approach. Using occupant counts as an exogenous input could further reduce the prediction error by 5%-6%. The findings of this paper could shed light on the plug load prediction for building control optimizations such as model-predictive control.
Article
Predicting building energy consumption is essential for planning and managing energy systems. In recent times, numerous studies focus on load forecasting models dealing with a wide range of different methods. In addition to Artificial Neural Networks (ANN), especially Support Vector Machines (SVM) have been studied. Various research work showed the success and superiority of ANN and SVM for load predictions, where frequently, SVM outperformed ANN models. In this study, data-driven thermal load forecasting performance of ε-SVM Regression (ε-SVM-R) based on a Radial Basis Function (RBF) and a polynomial kernel is compared to the outcome of two Nonlinear Autoregressive Exogenous Recurrent Neural Networks (NARX RNN) of different depths. For demonstration, historical data from a non-residential district in Germany is used for training and testing to predict monthly loads. The evaluation of the resulting predictions show that NARX RNNs yields higher accuracy than (ε-SVM-R) models, in combination with comparable computational effort.
Article
Many data-driven algorithms are being explored in the field of building energy performance estimation. Choosing an appropriate method for a specific case is critical to guarantee a successful energy operation management such as measurement and verification. Currently, little research work on assessment of different data-driven algorithms using real time measurement data sets is available. In this paper, five commonly used data-driven algorithms, ARX, SS, N4S, discretized variable BN and continuous variable BN, are used to estimate HVAC related electricity energy consumption in a university dormitory. In practice, total energy consumption data is easily accessible, while separated HVAC energy consumption data is not commonly available due to expensive sub-metering and/or the complexity of mechanical and electrical layouts. A virtual sub-meter based on a decomposition method is proposed to separate HVAC energy consumption from the total building energy consumption, which is derived from an improved Fourier series based decomposition method.