Acta Polytechnica CTU Proceedings


https://doi.org/10.14311/APP.2022.38.0090
Acta Polytechnica CTU Proceedings 38:90–96, 2022 © 2022 The Author(s). Licensed under a CC-BY 4.0 licence

Published by the Czech Technical University in Prague

ENERGY PERFORMANCE ESTIMATION FOR LARGE BUILDING
PORTFOLIOS WITH MACHINE LEARNING-BASED TECHNIQUES

Frédéric Monteta, ∗, Alessandro Pongellib, Jonathan Riala,
Stefanie Schwabc, Jean Henneberta, Thomas Jusselmeb

a University of Applied Science of Western Switzerland (HEIA-FR, HES-SO), iCoSys Institute, Bd. de Pérolles
80, 1705 Fribourg, Switzerland

b University of Applied Science of Western Switzerland (HEIA-FR, HES-SO), ENERGY Institute, Bd. de
Pérolles 80, 1705 Fribourg, Switzerland

c University of Applied Science of Western Switzerland (HEIA-FR, HES-SO), TRANSFORM Institute, Bd. de
Pérolles 80, 1705 Fribourg, Switzerland

∗ corresponding author: frederic.montet@hefr.ch

Abstract. Building operation is responsible for 28 % of the world’s carbon emissions. In this context,
establishing priorities in refurbishment strategies at the scale of a city or a group of buildings is
important. Such procedures are usually led by experts in energy performance and, therefore, they are
rarely carried out due to their long and costly nature.

This research aims at the estimation of building energy performance to pave the way towards
finding near-optimal refurbishment strategies. Thanks to the identification of easily-accessible building
characteristics, the method applies machine learning models to scan a building portfolio based on
a low level of details. The results show good potential to identify low-performer buildings with simple
machine learning methods. It also opens the door for further improvements through the inclusion of
supplementary building features at the input of the predictive system.

This work includes (a) the integration of a knowledge database thanks to the Swiss CECB energy
performance certificates, referencing more than 70 000 buildings, (b) the preparation of a training data
set through the selection of relevant physical characteristics of buildings (input) and the corresponding
energy consumption labels (output), (c) the development of predictive models used in a supervised way,
(d) their evaluation on an independent test set.

Keywords: Refurbishment strategies, machine learning, energy performance certificates.

1. Introduction
In a world where climate change imposes major adjust-
ments in order to slow down the rise in environmental
temperature, one important area to improve is the
planning of building renovations. Thanks to changing
construction techniques and equipment, building ren-
ovations became a major focus of energy strategies of
governments.

Switzerland, for example, is implementing a strat-
egy called “Energy Strategy 2050” which aims, among
other things, to reduce the energy consumption of
buildings through incentives [1]. This, in the coming
years, may cause a rush to renovate old buildings and
to prioritise the management of renovations due to
the costs and timeframes imposed. For building stock
managers it means having to priorities and create
a ranking of interventions that can lead to energy im-
provements, but which do not involve high costs. In
Switzerland there is the Cantonal Energy Certificate
for Buildings (CECB) [2], which allows the assessment
of the current state of the building and the planning
of a possible renovation. It allows to attribute an en-
ergy label to the efficiency of the building envelope,
which describes the quality of the thermal envelope

including roof, wall, floor and window insulation, and
also takes into account thermal bridges and the shape
of the building. A second label is given to the overall
energy efficiency including heat demand, electricity
demand, own production of electricity as well as the
building’s equipment for heat and domestic hot water.
The labels are divided into 7 classes: from A, the
best class, to G, the worst class compared to a refer-
ence. This certificate is compulsory in cantons such
as Geneva, Vaud, Fribourg, Neuchâtel, Nidwald, Zug
and Zürich, in case of sale and/or renovation of the
building. This has led to the creation of a database
with more than 70 000 certificates describing the phys-
ical characteristics and energy performance of each
building.

However, this certificate oblige experts to travel
to the building site in order to collect the various
data needed for in-depth analysis, increasing the cost
and time required to analyse a large building stock.
In recent years it has also become possible, through
portals such as Registre fédéral des bâtiments et des
logements (RegBL) [3], Cantonal geoportals [4, 5],
Google street view [6] and Google maps [7], to find
precise data or to collect it remotely. Therefore, in the

90

https://doi.org/10.14311/APP.2022.38.0090
https://creativecommons.org/licenses/by/4.0/
https://www.cvut.cz/en


vol. 38/2022 Energy performance estimation for large building portfolios . . .

Figure 1. Map of Switzerland showing the concentration of building energy performance certificates (CECB) per
cantons.

following chapters we will try to answer the main ques-
tion: “How can the available online data be used to
quickly classify a building stock energy performance?”.

In this paper, we investigate the use of Machine
Learning (ML) approaches as a solution to estimate
automatically the less performing buildings according
to the CECB methodology. The proposed method
is to take as input of the ML systems easy-to-find
building characteristics that do not need the interven-
tion of an energy expert. The output of the system is
a prediction of the CECB energy label of the building.
If functional, such approaches could provide a quick
and easy way to rank buildings by priorities of reno-
vations. The CECB association gave us access to his
data, under a data agreement for the protection and
sharing of sensitive data. We used these data to train
and test our ML systems.

The paper is organized as follows. Section 2 presents
the methods used to prepare the data and to select
and optimize the best performing models. Section 3
presents the obtained results according to specific met-
rics that we propose to evaluate their performances.
Finally, Sections 4 and 5 present discussions, conclu-
sions and future works.

2. Methods
This section introduces exploratory and preprocessing
phases performed on the data. Details regarding the
machine learning models selection, their optimization
and the process used to assess the results quality are
provided.

2.1. Data Exploration
As indicated in the previous Section 1, more than
70 000 certificates are in the dataset. To understand
the nature of the data, statistics were computed as

Figure 2. Evolution of the average heat transfer
coefficient in relation to the year of construction of
buildings.

a preliminary analysis. Knowing that in some cantons
the certificate is compulsory, it was verified that this
is reflected in the form of certificates in the database.
In Figure 1 it is possible to see how the dataset is
divided according to the different cantons and it is
possible to identify the cantons that have a compulsory
certificate.

A second analysis was carried out to verify data con-
sistency. In Figure 2, it is possible to see the average
heat transfer coefficient for each year. The evolution
of this value is getting lower through time, which cor-
relates with the increase in building insulation from
1970. This conclusion is similar to the findings of the
Energy and Renovation (eREN) project [8].

To get a better understanding of the subdivision
of the dataset, we then checked which categories are
present. They are subdivided according to the cat-

91



F. Montet, A. Pongelli, J. Rial et al. Acta Polytechnica CTU Proceedings

Categories CECB Qty. Percentage
Single-family building 40 590 55.10 %
Multi-family building 27 780 37.71 %
Administration 1 932 2.62 %
Mixed 1 802 2.45 %
School 1 374 1.87 %
Retail 79 0.11 %
Hotel 77 0.20 %
Restaurant 29 0.04 %

Table 1. Representation of the various building cate-
gories in the CECB dataset.

Original data After cleaning
A 1620 277
B 10695 4594
C 16036 10056
D 17271 10718
E 11615 7355
F 6950 4374
G 9474 4986

Table 2. Number of certificates per label before and
after cleaning the dataset.

egories of RegBL [3]. In the Table 1, it is possible
to see that the two most represented categories are
single-family building and multi-family building.

The subdivision in classes before the cleaning of the
dataset is present in the Table 2 where we can see
that the most represented class is D followed by other
classes and ended by the least represented A.

2.2. Data Preprocessing
The data received was in the form of several tables. To
use them, the latter were merged to combine the data
in a single structure. Subsequently, a data cleaning
was carried out. After manual checks, outliers were
identified and removed when samples were outside two
standard deviations from the mean, thus removing
5 % of the data. In addition, missing values were
removed when their percentages were more than 90 %
per column or more than 80 % per samples, i.e. per
certificate.

The result of this part of the method is presented in
Table 2 as a number of certificates after the cleaning.

For all following machine learning (ML) tasks, a
balanced dataset is preferred1. To achieve this, sub-
sampling can be used. Since the class A is clearly
underrepresented with its 277 samples, sub-sampling
the data would induce a high loss of samples. For
this reason, class A and B are merged before sub-
sampling for all further ML tasks. Moreover, as A
and B building classes are best performers, they are
not targeted by refurbishment plans.

1A dataset is said balanced when the number of samples per
class is equal.

2.3. Model Performance Exploration
The modelling method includes 3 steps.
(1.) Pre-processing routines were carried out. The

latter include an ordinal encoding of the data as
well as a normalization by computing the Z-score
on all variables.

(2.) Then, eighteen different models from the Scikit-
learn library were selected when appropriate for
a classification task and trained with a k-fold cross
validation with k = 3 [9].

(3.) The evaluation of all models is made based on
their F1-scores and accuracy metrics to select the
two most promising ones.

2.4. Optimization and Model Selection
On the two best models from the previous section,
optimization and training is performed to allow for
final model selection. Since the optimization tasks is
memory expensive, the research for optimal parame-
ters has been done on a reduced number of samples
per class selected in a random manner.

The optimization includes a cross-validated random-
ized search where k = 3 on a selection of parameters
(see Section 3). Model training is computed as in
step 3 from Section 2.3. Finally, the overall best
model is selected given its per-class metrics (Accuracy
and F1) and confusion matrix.

2.5. Special Cases Identification
As a last step of the method, a more in-depth analysis
of special cases is performed to identify the reasons
behind a high prediction error.

A distance between classes is computed with
a method taken from ordinal regression problems [10].
To compute this distance, let y = A and ŷ = G
be a sample’s class and its estimation. Their en-
coded version would be y′ = (1, 0, 0, 0, 0, 0, 0) and
ŷ′ = (1, 1, 1, 1, 1, 1, 1). The distance d between the two
classes can then be calculated with d =

∑n
i=1 |y

′
i − ŷ

′
i|.

Once obtained, all samples were ordered by descend-
ing order from their distance d.

3. Results
3.1. Model Performance Exploration
The eighteen algorithms were trained on a resampled
dataset to make it balanced. Performances obtained in
the Table 3 allow for the comparison of all algorithms
in order to identify the two best candidates for energy
performance certificates predictions.

The DummyClassifier performs a classification in
a random manner with 0.17 accuracy. The latter
result sets a baseline above which an algorithm learns
the characteristics of the data. Since most algorithms
performed with a F1-score and accuracy between 0.3
and 0.5, the problem at hand can be characterized as
difficult.

92



vol. 38/2022 Energy performance estimation for large building portfolios . . .

Model Accuracy F1-score
HistGradientBoostingClassifier 0.50 0.49
RandomForestClassifier 0.50 0.48
ExtraTreesClassifier 0.49 0.48
BaggingClassifier 0.49 0.48
GradientBoostingClassifier 0.49 0.47
MLPClassifier 0.47 0.46
SVC 0.46 0.45
NuSVC 0.46 0.45
LogisticRegression 0.45 0.43
DecisionTreeClassifier 0.43 0.43
AdaBoostClassifier 0.43 0.42
LinearDiscriminantAnalysis 0.43 0.41
ExtraTreeClassifier 0.39 0.38
KNeighborsClassifier 0.39 0.38
LinearSVC 0.41 0.37
SGDClassifier 0.40 0.35
RidgeClassifier 0.39 0.34
NearestCentroid 0.35 0.34
Perceptron 0.33 0.32
PassiveAggressiveClassifier 0.32 0.31
BernoulliNB 0.32 0.31
GaussianNB 0.30 0.30
DummyClassifier 0.17 0.05

Table 3. Model performances exploration. Models
are ordered by F1-scores and then, by their accuracy
metrics.

Before After Gain
Catboost 0.59 - -
HistGradientBoosting 0.49 0.58 18 %
RandomForest 0.48 0.57 19 %

Table 4. Comparison of the best classifiers with their
F1-score average across class.

The top 5 algorithms include recent classi-
fication models performing with similar result
around 0.48 F1-score. The MLPClassifier2 per-
formed with moderately good result. The best
selected candidates for the next step of the
method are the HistGradientBoostingClassifier
and RandomForestClassifier.

3.2. Optimization and Model Selection
The fine tuning procedure was performed
on HistGradientBoostingClassifier and
RandomForestClassifier models. In addition,
one state-of-the-art gradient boosted tree implemen-
tation was added from the Catboost library, without
fine-tuning [11].

As Table 4 shows, all models are reaching perfor-
mances above or close to 0.5 without optimization,
thus already allowing for an educated guess. When
parameters from Listing 1 are used, substantial gains
can be obtain within the order of ∼20 %; thus proving
the value of randomized search.

2Stands for Multi-layer Perceptron, i.e. a neural network
with the default parameters from the Scikit-learn library.

b e s t _ p a r a m s _ h i s t _ g r a d i e n t _ b o o s t i n g = {
’ l 2 _ r e g u l a r i z a t i o n ’ : 0 . 0 0 0 8 1 ,
’ l e a r n i n g _ r a t e ’ : 0 . 1 1 8 8 9 ,
’ max_bins ’ : 6 0 ,
’ max_leaf_nodes ’ : 1 1 ,
’ m i n _ s a m p l e s _ l e a f ’ : 87

}

best_params_random_forest = {
’ n _ e s t i m a t o r s ’ : 8 0 0 ,
’ m i n _ s a m p l e s _ s p l i t ’ : 5 ,
’ m i n _ s a m p l e s _ l e a f ’ : 1 ,
’ m a x _ f e a t u r e s ’ : ’ s q r t ’ ,
’ max_depth ’ : 8 0 ,
’ b o o t s t r a p ’ : F a l s e

}

Listing 1. Best parameters.

Unbalanced Balanced
AB 0.76 0.82
C 0.69 0.62
D 0.60 0.54
E 0.45 0.44
F 0.31 0.47
G 0.68 0.67
Accuracy 0.60 0.59
F1 avg. 0.59 0.59

Table 5. Global building efficiency F1-scores.

A final F1-score average across classes of 0.59 makes
the Catboost based model the more performant of
the selection. Also, both algorithms from Scikit-learn
library had a notably high performance gain, making
them reach a scores comparable to Catboost.

3.3. Best Model Performances
The Catboost model, using gradient boosting on de-
cision trees, is the final model selected. This section
introduces it in three phases. First, with explaining
an additional data processing task. Second, by pre-
senting the model performances in depth. Finally, by
analyzing the errors produced by the model.

3.3.1. Model Performance
The final model performances are summarized in
Table 5. In average, the model reaches an accu-
racy of ∼0.6. This represents an improvements of
∼350 % compared to the baseline of 0.17 given by the
DummyClassifier from Section 3.

On unbalanced data, scores have a great variability,
which makes the model quality hard to evaluate with
confidence. Since unbalanced data has generally more
samples, the awaited behavior is a higher score, but
this isn’t the case for all classes. A possible expla-
nation lies in the selection of more representatives
samples while sub-sampling.

93



F. Montet, A. Pongelli, J. Rial et al. Acta Polytechnica CTU Proceedings

(a). Unbalanced dataset. (b). Balanced dataset.

Figure 3. Confusion matrices for global efficiency.

On balanced data, the comparison between classes
gives an insight on the difficulty to evaluate each class.
From AB to F , scores are decreasing from 0.82 to 0.47.
This shows that the lower the efficiency of a building
is, the harder it becomes to predict its class. Both
end of the class spectrum – AB and G – have higher
scores. This behavior is probably due to the absence
of strict intervals i.e. a certificate made on a highly
inefficient building would have a class higher than G,
but since it doesn’t exist, it is classed as G.

More details about the predictions uncertainty are
given by the confusion matrices in Figure 3 where
the precision of the models are presented visually.
A highlighted diagonal is predominant in both the
unbalanced and balanced plots; showing the correct
predictions. On each side of the diagonal, the predic-
tions made with an error of one class above or below
the true label are represented.

3.3.2. Prediction Reliability
To evaluate how the false predictions are spread
around the diagonal, the plot on Figure 4 shows the
cumulative density of the distances between predicted
and true value. The accuracy of ∼60 % is visible where
the distance is 0. Then, in case of a wrong prediction,
there is a ∼90 % probability that the true class is only
a letter away; ∼100 % for two letters, and so on. This
highlights that wrong predictions are generally not
far away from their target.

3.4. Online variables
To assess whether the necessary input variables are
available online to speed up the classification process
with ML-based techniques, a wide search was carried
out on the various portals listed in the Section 1. The
Figure 5 presents the most discriminant variables in
the Catboost algorithm. For brevity, only the ten
firsts are presented.

Of the most important variables for the operation
of the algorithm, we can easily find the year or era of
construction of the building on the RegBL site through

Figure 4. Cumulative distribution of the distances
between true and predicted global efficiency given two
two different metrics.

the interactive building map. A second variable that
can easily be found is the building width through
Google Maps for example as the measurement tool
can be used to obtain a value of the desired building.

A third and a fourth that can be found on the
internet, but is subject to restrictions, are the year
of construction of the energy agent and the energy
source. They are available in the RegBL database, but
this type of data is only granted after authorization
and verification of credentials.

All other variables are not accessible as they are
not present on any platform at the moment.

4. Discussion
The applicability of the methods depends on
(1.) the accuracy of the model that is used to predict

performance classes, and
(2.) the easy online access of inputs of the predicting

model describing the physical properties of buildings
to be classified.

94



vol. 38/2022 Energy performance estimation for large building portfolios . . .

Figure 5. Feature importance of the first ten variables used by Catboost.

Regarding (1.), the Catboost model is promising as it
has close to 60 % accuracy. Moreover, the prediction
is only 1 class away in 90 % of the buildings. This
seems to be highly acceptable if the method decreases
dramatically the time spent to assess the building
performance.

In order to be able to understand how to further im-
prove prediction, it is necessary to start by analysing
the process used step by step, starting with the collec-
tion of data for the generation of the CECB certificate.

The initial dataset is the result of certificates exe-
cuted by many CECB experts. Several points in the
creation of a CECB certificate are decided through
visual inspection and thus based on the knowledge of
the individual expert. For example, in order to get
a U-value for walls, the expert can enter a proposal for
the composition of the wall by visual inspection and
the program calculates a U-value based on what the
expert insert. This means that each expert, according
to his knowledge, can enter or omit data, but in the
end the program still manages to generate a complete
certificate. This, for example, explains the missing
data for some certificates. In order to have a more
complete dataset, it would be important to reduce
the possibility of omitting data for the generation of
a certificate to a minimum.

Regarding the type of certificate, before 2012, a sim-
pler type of certificate with fewer parameters was in
place, then a more detailed certificate was adopted.
This certainly causes a lack of data on some of the
certificates used. Filling the gaps and convert less de-
tailed certificates into more detailed ones by entering
the missing data could be a solution to improve the
base set.

During the data cleaning, arbitrary choices were
made to eliminate values with the 3 sigma method,
to eliminate columns with 90 % of the data missing as
well as certificates with 80 % of the data missing. This

system can clearly decrease the accuracy by removing
variables and especially population from the data
set. A more thorough analysis for cleaning the data
should be done, so as to be sure that all inconsistent
values are removed and that all consistent values are
kept. The same can be done by keeping variables
considered important as well as certificates to increase
the population.

Finally, when it comes to prediction, changing the
settings of the algorithm could lead to further im-
provement of the dataset. A better tuning can lead
to a more precise prediction.

Moving on to the second point (2.), it was identified
in Section 3.4 that only two variables are easily acces-
sible. In the next part of the article we will analyse
each variable in the Figure 5, exploring possible meth-
ods of providing representative values of the analysed
building for the data required.

In order, the most important variable is the heat
transfer coefficient of walls, which cannot be found
online. One solution would be to send a person on
site to do an analysis in order to give a value to this
variable. A second possibility that should be explored
is to use artificial intelligence to reconstruct this value,
for example from an image of the wall of the building
and the year of construction, so that it can recognize
some key features and then give a value based on
other similar buildings.

Moving on to the second variable in the list is the
Energy source. As already discussed in the Section 3.4,
it can be found on the RegBL website with restrictions.
However, this data may be in the possession of the
owners or managers of the building and could therefore
be entered easily.

The year of construction of the heating, as men-
tioned in the Section 3.4, can be found on the RegBL
website with restrictions. Also this variable should be
checked if the owner or manager of the building is in

95



F. Montet, A. Pongelli, J. Rial et al. Acta Polytechnica CTU Proceedings

possession of it.
The envelope coefficient is not found directly but

could be calculated from other variables. Online it is
possible to measure the building perimeter through
Google Maps. On the RegBL website it is possible
to find the number of floors in the building and the
ground area. Using for example an average value for
the height of a floor, the various parameters can be
combined. Multiplying the perimeter by the num-
ber of floors and the average height and then adding
the floor area twice gives an approximation of the
envelope surface. Dividing the envelope surface by
the number of floors multiplied by the floor surface
gives an approximate value of the envelope coefficient.
Clearly an approximate value that should be checked
for potential and especially for possible errors brought
with it.

The utilisation rate of heat generator for hot water
cannot be found online and so there are two possibili-
ties to find this data. The first is that the manager or
owner of the building knows this value. The second is
to send an expert on site to make an assessment.

The energy reference surface can also be estimated
from other variables. On the RegBL website we find
the floor area of the building and the number of floors,
multiplying these variables together we find a rough
estimate. Clearly, the reliability of predicting the
correct energy class using this approximation must be
checked.

The climate station is not available online, but
knowing the address of the building it is possible to
indicate which weather station is to be used for the
calculation.

The linear thermal bridge building base lWF is like
the heat transfer coefficient of walls, is not available
online and it is possible to use the same solution
proposed.

It must be said that all missing variables could
be estimated or found easily with an expert on site.
Having only 10 variables to find would simplify and
speed up the work to be done on site.

5. Conclusion
In this paper we have highlighted the preliminary
reliability results of using a classification algorithm to
analyse a building. The result of ranking the building
in the good class with 60 % accuracy is a promising
result for future developments. It should be noted that
there is a 90 % probability of being in the adjacent
class, which brings value to the work done.

In addition, the work carried out to check the online
presence of the most important variables for prediction
has shown that it is still premature to find the exact

value online. Nevertheless, it may be possible to
recreate some them by developing further techniques.

Some recommendations for future work are neces-
sary. For example, A special attention must be taken
when merging the different data available during the
preparation phase, as this could lead to consecutive
errors in the other phases of the work. A special atten-
tion must also be taken during the cleaning phase to
ensure that the maximum amount of data is available
to carry out the work.

Acknowledgements
The authors would like to express their gratitude to Ms
Karine Wesselmann and the CECB association for provid-
ing the data for this work.

They would also like to thank Professor Mylène Devaux
and the iTEC institute for their collaboration.

Financial support is gratefully acknowledged from the
HEIA-FR Smart Living Lab research program.

References
[1] Swiss Federal Office of Energy. Stratégie énergétique

2050, 2020. [2021-11-05]. https://www.bfe.admin.ch/
bfe/fr/home/politik/energiestrategie-2050.html

[2] Association GEAK-CECB-CECE. Le Certificat
énergétique cantonal des bâtiments (CECB).
[2021-11-05]. https://www.cecb.ch/

[3] Office fédéral de la statistique. Registre fédéral des
bâtiments et des logements (RegBL). [2021-11-05].
https://www.housing-stat.ch/fr/index.html

[4] Canton de Fribourg. Portail cartographique du canton
de Fribourg. [2021-11-05]. https://map.geo.fr.ch/

[5] Canton de Neuchâtel. Portail cartographique du canton
de Neuchâtel. [2021-11-05]. https://sitn.ne.ch/

[6] Google LLC. Google Street View. [2021-11-05].
https://www.google.com/intl/en_ch/streetview/

[7] Google LLC. Google Maps. [2021-11-05].
https://www.google.ch/maps/

[8] S. Schwab, L. Rinquet, M. Devaux, et al. Rénovation
énergétique. Tech. rep., Fribourg, Suisse, 2018.

[9] F. Pedregosa, G. Varoquaux, A. Gramfort, et al.
Scikit-learn: Machine learning in Python. Journal of
Machine Learning Research 12:2825–2830, 2011.

[10] J. Cheng, Z. Wang, G. Pollastri. A neural network
approach to ordinal regression. In 2008 IEEE
International Joint Conference on Neural Networks
(IEEE World Congress on Computational Intelligence),
pp. 1279–1284. 2008.
https://doi.org/10.1109/IJCNN.2008.4633963

[11] L. Prokhorenkova, G. Gusev, A. Vorobev, et al.
CatBoost: unbiased boosting with categorical features,
2019. arXiv:1706.09516

96

https://www.bfe.admin.ch/bfe/fr/home/politik/energiestrategie-2050.html
https://www.bfe.admin.ch/bfe/fr/home/politik/energiestrategie-2050.html
https://www.cecb.ch/
https://www.housing-stat.ch/fr/index.html
https://map.geo.fr.ch/
https://sitn.ne.ch/
https://www.google.com/intl/en_ch/streetview/
https://www.google.ch/maps/
https://doi.org/10.1109/IJCNN.2008.4633963
http://arxiv.org/abs/1706.09516

	Acta Polytechnica CTU Proceedings 38:90–96, 2022
	1 Introduction
	2 Methods
	2.1 Data Exploration
	2.2 Data Preprocessing
	2.3 Model Performance Exploration
	2.4 Optimization and Model Selection
	2.5 Special Cases Identification

	3 Results
	3.1 Model Performance Exploration
	3.2 Optimization and Model Selection
	3.3 Best Model Performances
	3.3.1 Model Performance
	3.3.2 Prediction Reliability

	3.4 Online variables

	4 Discussion
	5 Conclusion
	Acknowledgements
	References