Deep Learning Approach based on FCRBM for Optimization
of Electric Energy Production
Chaimaa Fouhad, Mohamed El Khaili, Mohammed Qbadou
ENSET Mohammedia, Hassan II University of Casablanca, BP 159, Mohammedia, Morocco
Keywords: Deep Learning, Energy production, Artificial Neural Network, Conditional Restricted Boltzmann Machine
CRBM, Factored CRBM.
Abstract: This correspondence features Deep Learning's commitment to the electrical energy sector. An overview of
the concept will highlight the commitment of this innovation in enhancing the creation of electrical energy,
prior to setting out on the decision of the model from which we will begin. Latter's choice is made after
comparative studies between the different models used by other authors in their previous publications on the
subject.
This investigation in the end drove us to embrace the Factored Conditional Restricted Boltzmann Machine
"FCRBM" model as a model that was viewed as powerful in correlation with others in a similar class. The
FCRBM is a five-layer model, including three layers of the CRBM strategy, to which two new added
substance layers have been added to work on the exactness and give new usefulness and functionality.
1 INTRODUCTION
The creation of electrical energy has been a subject
of worry for scientists and researchers since the
innovation of the electricity. Certainly, production
without losses doesn't exist. These energy losses
have offered way to the topic of how to enhance the
production of electricity.
Since the introduction of mechanical methods
has not been agreeable, the commitment of new is
exceptionally attractive. Deep learning is one of the
technologies that can provide the solution for the
problem. Profound Learning is broadly used to
enhance the creation of electrical energy because of
its different techniques and errand computerization.
We will look at the different topics related to our
context to help us choose the most appropriate and
effective method.
This correspondence comprises of four segments.
The main presents a concise best in class.
Conditional Restricted Boltzmann Machine (CRBM)
and Factored Conditional Restricted Boltzmann
Machine (FCRBM) are depicted in the subsequent
segment. The third part presents our model with its
execution engineering. Reproduction results will be
accessible once the datasets are gotten.
2 RELATED WORKS
A significant element of future power matrices is
forecasting energy utilization throughout a wide
scope of time horizons; subsequently, expect total
interest as well as to extend the individual structure
so that circulated creation assets can be sent by the
nearby utilization, particularly because of the huge
gadgets (Marhoum et al., 2021).
Furthermore, the interest decay makes it
conceivable to investigate energy utilization designs
and distinguish energy protection goals. Likewise,
determining transient energy utilization permits
directors to design energy utilization over the long
run, move energy utilization to off-top periods, and
get ready more good energy buy plans. Generally,
demand forecasting can be considered to fall into
three categories:
Short-term forecasts are generally applied at
intervals of one hour to one week,
Medium-term forecasts are generally one
week to one year,
Moreover, long-term forecasts are for more
than one year.
Energy forecasting an unpredictable issue since
it relies upon the intricacy of the structure's energy
conduct and the vulnerability of the affecting
variables, prompting incessant changes popular.
Fouhad, C., El Khaili, M. and Qbadou, M.
Deep Learning Approach based on FCRBM for Optimization of Electric Energy Production.
DOI: 10.5220/0010733900003101
In Proceedings of the 2nd International Conference on Big Data, Modelling and Machine Learning (BML 2021), pages 345-349
ISBN: 978-989-758-559-3
Copyright
c
2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
345
These variances are because of the structure
engineering and warm properties of the actual
materials utilized, tenants and conduct, climatic
conditions, and sub-level framework segments like
lighting or HVAC (warming, ventilation and
cooling). (Yang et al., 2014)
A decent forecast is additionally the aftereffect
of viable prescient investigation. It is in this setting
that few knights of science have given themselves
the test to acquire a successful prescient model
dependent on the commitment of new innovations.
Amir Mosavi did an examination dependent on
Machine Learning and showed how the methods
utilized by ML could work on the precision in the
forecast of the creation of electrical energy. On a line
close to Amir Mosavi, Marijana Zeki'c-Suˇsac
completed an examination on the commitment of ML
in the creation of energy in the public area and
exhibited the advantages that this new innovation
could bring to the electrical energy area. (Cost et al.,
2013)
Elena (Mocanu et al.,2016) put out an objective
of working on the productivity of the whole
electrical framework and exhibited that by utilizing
neural organization-based forecast strategies,
learning procedures are relied upon to work on the
presentation and precision of expectation by
permitting more elevated levels of reflection. To
accomplish this, two stochastic models were
contemplated: the Conditional Restricted Boltzmann
Machine (CRBM) and Factored Conditional
Restricted Boltzmann Machine (FCRBM). A
correlation showed that for the issue of anticipating
energy creation, the FCRBM outperformed the
Artificial Neural Network (ANN), Support Vector
Machine (SVM), Recurrent Neural Networks (RNN)
and the CRBM.
Figure 1: The overall design of Conditional Restricted
Boltzmann Machines, where u is the contingent history
layer (input), "h" is the secret layer, and "v" is the apparent
layer (yield), where means twofold neurons, addresses the
genuine qualities, and the other sort of units are Gaussian
neurons.
3 DESCRIPTION OF FACTORED
CONDITIONAL RESTRICTED
BOLTZMANN MACHINE
Taylor and Hinton presented the Factored Condition
Restricted Boltzmann Machine (FCRBM), where
they add styles and the idea of figured,
multiplicative, three-way associations to anticipate
various styles of human motion.
Figure 2: The overall design of Factored Conditional
Restricted Boltzmann Machines, where "u" is the
contingent history layer (input), h is the secret layer, y is
the style layer, and "v" is the noticeable layer (yield),
where indicates double neurons, addresses the genuine
qualities and the others are Gaussian qualities.
Initially, FCRBMs comprise of the past three
layers from CRBM and two new presented layers for
styles and highlights. Nonetheless, to meet our
requirements, we decreased the style and the
highlights layers to one, and we utilized it to address
various boundaries valuable for expectation.
All the more decisively, after the decrease as
referenced above, FCRBM comprises of:
(1) A genuine esteemed apparent layer “v”.
(2) A genuine esteemed history layer "u"
(i.e.,
𝑢
:
, where 𝑁∈ℕ).
(3) A double secret layer "h".
(4) A style layer “y”.
Every one of the above layers is fundamental for
the accomplishment of the FCRBMs. The noticeable
layer encodes the current upsides of a period series
which should be anticipated. The historical backdrop
of the time succession, being the premise of such
forecasts, is encoded on the set of experiences layer.
The secret layer ensures the disclosure of significant
highlights fundamental for investigating the time
succession, while the style layer encodes various
BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)
346
boundaries helpful in the expectation. To gain
proficiency with the intrinsic relations between these
layers, undirected or coordinated loads and factors,
as displayed in Figure 2, are utilized as associations.
All the more officially, FCRBM characterizes a
joint likelihood conveyance over the apparent "v",
and covered up "h", neurons. The joint conveyance
is adapted on the past N perceptions "u", model
boundaries (∙∙) (i.e., 𝑊
,
𝑊
,𝑊
,𝐴
,𝐴
,𝐴
,𝐵
,𝐵
,𝐵
), and the style layer
"y". Like CRBM, it is expected parallel stochastic
secret units and genuine esteemed noticeable units
with added substance, Gaussian commotion. For
notational ease, as in the first paper [20], we assume
σi = 1.
The complete energy work for this model is:
𝐸𝑣
𝑎ℎ
𝑏
𝑣
𝑤
∘𝑦
𝑤
∘ℎ
𝑤
(1
)
Where 𝑊
, 𝑊
, and 𝑊
, address the noticeable
figured, covered up considered, and name loads
calculated, individually. X ◦ Y is the Hadamard
item, otherwise called the component insightful
item, between networks X and Y, and f addresses the
records, all things considered.
The terms 𝑎 and 𝑏
are called dynamic biases and
are defined as:
𝑎𝑎𝐴
𝑢
𝐴
∘𝑦
𝐴
(2
)
𝑏
𝑏𝐵
𝑢
𝐵
∘𝑦
𝐵
Where 𝐴
, 𝐴
, 𝐴
, 𝐵
, 𝐵
, 𝐵
, are the
associations from the comparing layer to the
variables. Just as the load’s associations are free
boundaries that should be prepared as itemized in
the accompanying area.
3.1 Inference in FCRBMs
In FCRBMs, probabilistic induction implies
deciding two contingent circulations. The first is the
likelihood of the secret layer adapted on the wide
range of various layers (i.e., p (h|v, u, y)), while the
second is the likelihood of the current layer molded
on the others (i.e., p (v|h, u, y)). Since there are no
associations between the neurons in a similar layer,
deduction can be made in equal for every unit type,
prompting:
𝑝
ℎ1
|
𝒖,𝒗,𝒚
𝑠𝑖𝑔𝑏
𝑤
𝑦
𝑤
∘
𝑣
𝑤
(3
)
Where
𝑠𝑖𝑔𝑥 1/1 exp 𝑥
, and
𝑝
𝑣
|
𝒉,𝒖,𝒚
ℵ𝑎𝑤
𝑦
𝑤
∘
ℎ
𝑤
,𝜎
(4
)
Where for accommodation σ is picked to be 1.
3.2 Learning & Update Rules for
FCRBMs
The free parameters model (i.e., dynamical biases
and weights) are picked up utilizing Contrastive
Divergence, discussed previously in the previous
section, as displayed in Algorithm 1. The update
rules for each of these connections can be computed
by deriving the energy function in relation to each of
the variables. For more details on these derivations,
please refer to [18]. After calculating the derivatives,
the following update rules are found:
𝑤
𝑤
𝛼⟨ℎ
𝑣
𝑤
∘
𝑦
𝑤
⟩
⟨
ℎ
𝑣
𝑤
∘
𝑦
𝑤
⟩
(5
)
𝑤
𝑤
𝛼⟨𝑦
𝑣
𝑤
∘
ℎ
𝑤
⟩
𝑦
𝑣
𝑤
∘
ℎ
𝑤
𝑤
𝑤
𝛼⟨𝑣
𝑦
𝑤
∘
ℎ
𝑤
⟩
𝑣
𝑦
𝑤
∘
ℎ
𝑤
and the dynamic biases updates are:
𝐴
𝜏1
𝑢
𝐴
𝜏
𝑢
𝛼⟨𝑢
𝑦
𝑇
𝐴
𝑦
∘
𝑣
𝑇
𝐴
𝑣
⟩
𝑑𝑎𝑡𝑎
⟨
𝑢
𝑦
𝑇
𝐴
𝑦
∘
𝑣
𝑇
𝐴
𝑣
⟩
𝑟𝑒𝑐𝑜𝑛
(6
)
𝐴
𝜏1
𝑣
𝐴
𝜏
𝑣
𝛼⟨𝑣
𝑦
𝑇
𝐴
𝑦
∘
𝑢
𝑇
𝐴
𝑢
⟩
𝑑𝑎𝑡𝑎
⟨
𝑣
𝑦
𝑇
𝐴
𝑦
∘
𝑢
𝑇
𝐴
𝑢
⟩
𝑟𝑒𝑐𝑜𝑛
𝐴
𝜏1
𝑦
𝐴
𝜏
𝑦
𝛼⟨𝑦
𝑢
𝑇
𝐴
𝑢
∘
𝑣
𝑇
𝐴
𝑣
⟩
𝑑𝑎𝑡𝑎
⟨
𝑦
𝑢
𝑇
𝐴
𝑢
∘
𝑣
𝑇
𝐴
𝑣
⟩
𝑟𝑒𝑐𝑜𝑛
𝐵
𝜏1
𝑢
𝐵
𝜏
𝑢
𝛼⟨𝑢
𝑦
𝑇
𝐵
𝑦
∘
ℎ
𝑇
𝐵
ℎ
⟩
𝑑𝑎𝑡𝑎
⟨
𝑢
𝑦
𝑇
𝐵
𝑦
∘
ℎ
𝑇
𝐵
ℎ
⟩
𝑟𝑒𝑐𝑜𝑛
(7
)
𝐵
𝜏1
ℎ
𝐵
𝜏
ℎ
𝛼⟨ℎ
𝑦
𝑇
𝐵
𝑦
∘
𝑢
𝑇
𝐵
𝑢
⟩
𝑑𝑎𝑡𝑎
⟨
ℎ
𝑦
𝑇
𝐵
𝑦
∘
𝑢
𝑇
𝐵
𝑢
⟩
𝑟𝑒𝑐𝑜𝑛
𝐵
𝜏1
𝑦
𝐵
𝜏
𝑦
𝛼⟨𝑦
𝑢
𝑇
𝐵
𝑢
∘
ℎ
𝑇
𝐵
ℎ
⟩
𝑑𝑎𝑡𝑎
⟨
𝑦
𝑢
𝑇
𝐵
𝑢
∘
ℎ
𝑇
𝐵
ℎ
⟩
𝑟𝑒𝑐𝑜𝑛
𝑎
𝜏
1
𝑎
𝜏
𝛼
⟨
𝑣
⟩
𝑑𝑎𝑡𝑎
⟨
𝑣
⟩
𝑟𝑒𝑐𝑜𝑛
(8
)
𝑏
𝜏1
𝑏
𝜏
𝛼
⟨
ℎ
⟩
𝑑𝑎𝑡𝑎
⟨
ℎ
⟩
𝑟𝑒𝑐𝑜𝑛
Deep Learning Approach based on FCRBM for Optimization of Electric Energy Production
347
Algorithm 1: FCRBM training procedure
4 OUR APPROACH
Based on the results of the various articles,
particularly those mentioned above, for all that is
predictive analysis to improve the production of
energy in electrical systems, we will start with the
FCRBM model approach with some exceptions
related to our vision. Simulation results will be
available once the datasets are received.
4.1 Project Context
Deep Learning is a set of methods for machine
learning, the aim of which is to model data at a high
level using non-linear transformation architectures.
The aim is to take advantage in the field of energy
production of methods related to Deep Learning to
obtain an expected optimisation of energy
production. We will be particularly interested in the
transport layer for said energy improvement.
4.2 Modelling Aspect
The modelling process aims to obtain an
understandable result by the computer system. The
final solution is a series of iterations. Several steps
are necessary for this purpose. These successive
steps allow to refinement of the level of details of
the system to be carried out. Early steps provide
vision to very large grains and advance
understanding of the problem. In the current
environment, the choice is based on the FCRBM
model.
5 CONCLUSION
Energy forecasting is a troublesome issue since it
relies on the complexity of the building’s energy
behavior and the uncertainty of the influencing
factors, prompting incessant vacillations sought
after. We embraced the Factored Conditional
Restricted Boltzmann Machine "FCRBM" model
since it was viable contrasted with the others in a
similar class. It can withstand variances because of
building engineering and warm properties of the
actual materials utilized inhabitants and their
conduct, climatic conditions and sub-level
framework parts like lighting or HVAC (warming,
ventilation, and cooling). The useful tests will
approve the viability of our methodology in
correlation with comparable methods.
REFERENCES
Elena Mocanu, Phuong H. Nguyen, Madeleine Gibescu,
Wil L. Kling, Title: Deep learning for estimating
building energy consumption, Sustainable Energy,
Grids and Networks Volume 6, June 2016, Pages 91-9.
R. Marhoum, C. Fouhad, M. El Khaili, H. Ouajji,
Contribution to Multi-Energy Flow Management for
Building Energy Hub, Journal of Ubiquitous Systems
and Pervasive Networks, Volume 15, No. 1 pp. 27-34,
March 2021, DOI: 10.5383/JUSPN.15.01.004
A. Costa, M.M. Keane, J.I. Torrens, E. Corry, Building
operation and energy performance: Monitoring,
analysis and optimization toolkit, Appl. Energy 101
(2013) 310–316.
http://dx.doi.org/10.1016/j.apenergy.2011.10.037.
L. Yang, H. Yan, J.C. Lam, Thermal comfort and building
energy consumption implications—A review, Appl.
Energy 115 (2014) 164–173.
http://dx.doi.org/10.1016/j.apenergy.2013.10.062.
M. Rafik, A. Bahnasse, A. Khiat, O. Bouattane, H. Ouajji,
M. El khaili, Towards a Smart Energy Sharing in
Micro Smart Grid Adopting SDN Approach, Procedia
Computer Science,2019, Pages 717-724, ISSN 1877-
0509, https://doi.org/10.1016/j.procs.2019.04.095
M. Simoes, R. Roche, E. Kyriakides, S. Suryanarayanan,
B. Blunier, K. McBee, P. Nguyen, P. Ribeiro, A.
Miraoui, A comparison of smart grid technologies and
progress in Europe and the US, IEEE Trans. Ind. Appl.
48 (4) (2012) 1154–1162.
http://dx.doi.org/10.1109/TIA.2012.2199730.
E.A. Bakirtzis, C.K. Simoglou, P.N. Biskas, D.P. Labridis,
A.G. Bakirtzis, Comparison of advanced power
system operations models for large-scale renewable
integration, Electr. Power Syst. Res. 128 (2015) 90–
99. http://dx.doi.org/10.1016/j.epsr.2015.06.025.
BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)
348
M. Krarti, Energy Audit of Building Systems: An
Engineering Approach, seconded., Mechanical and
Aerospace Engineering Series, 2011.
A. Foucquier, S. Robert, F. Suard, L. Stéphan, A. Jay,
State of the art in building modelling and energy
performances prediction: A review, Renewable
Sustainable Energy Rev. 23 (2013) 272–288.
http://dx.doi.org/10.1016/j.rser.2013.03.004.
C.-C. Chang, C.-J. Lin, Libsvm: A library for support
vector machines, ACM Trans. Intell. Syst. Technol. 2
(3) (2011) 27:1–27:27.
M. Lukoševičius, H. Jaeger, Reservoir computing
approaches to recurrent neural network training,
Comput. Sci. Rev. 3 (3) (2009) 127–149.
M. Lukoševičius, A practical guide to applying echo state
networks, in: Neural Networks: Tricks of the Trade,
in: Lecture Notes in Computer Science, vol. 7700,
2012.
Deep Learning Approach based on FCRBM for Optimization of Electric Energy Production
349
