Civil Engineering Reference
In-Depth Information
(
)
−
.
(
−
)
−
.
(
−
)
−
.
(
−
)
y
k
0
9417
y
k
1
0
2779
y
k
2
0
008656
y
k
3
+
.
(
−
)
+
.
(
−
)
+
.
(
−
)
0
06594
y
k
4
0
1009
y
k
5
0
06291
y
k
6
=
0
.
0001126
u
(
k
)
+
v
(
k
).
(4.17)
4.2.2 Artificial Neural Network Model
The indoor climate generated inside a room has a nonlinear nature, and taking into
account the results provided by the linear models described in Sect.
4.2.1
, an ANN
can be selected to obtain a nonlinear model structure determined by input-output
data in such a way that the results estimated by the model supply a better adjustment
to the ones measured inside the room.
An ANN is a calculation element inspired on the neurons network of the nervous
system of any living organism. It is composed of elements (neurons or nodes) con-
nected in parallel and whose joint action is able to reproduce complex functions.
Moreover, the existing connections among nodes are adaptable, and thus, they allow
to modify the global function of the network (Arahal et al.
2006
). As mentioned
allows to approximate any nonlinear static mapping between input and output vari-
ables provided that certain conditions are met. A wide description of the different
kinds of ANNand themethodology used for their training and validation can be found
in Arahal et al. (
2006
). Moreover, static ANN can be used to represent the dynamic
behaviour of a system by being fed with delayed values of the variables (Arahal et al.
1998
).
Therefore, in this section, a room-level model able to represent accurately the
dynamic behaviour of the main environmental variables that influence users' comfort
has been developed. More specifically, it has been considered that these variables are
indoor air temperature, indoor air relative humidity and the indoor CO
2
concentration,
and thus, the room-level model is composed of three submodels. The particularities
associated with each of the submodels, as the number of inputs, are explained in
subsequent sections. Hence, the steps followed to obtain each of the previous models
are:
•
To determine the ANN structure and the number of hidden layers
. To obtain
a nonlinear smooth map it is only necessary a hidden layer (Huang et al.
2000
).
Hence, the ANN obtained for the room-level model is a multilayer perceptron
composed of a hidden layer of
N
h
nodes with a sigmoidal activation function
and one output node with linear characteristic. In addition, it has been combined
with Tapped Delay Lines (TDL) blocks which are used to provide an appropriate
number of past values for the inputs (Arahal et al.
1998
).
•
Selection of suitable data sets in order to train the ANN
. The second step is
related to the selection of appropriate data sets which are used for the training
process of the ANN. Another important point, which is strongly connected with
the previous one, is the choice of the variables necessary to estimate each of the
