Subsystems and Disturbance Models - Comfort Control in Buildings - page 92

submodels that compose the room-level model: indoor air temperature, indoor
relative humidity and the indoor CO 2 concentration and their associated ranges.
Moreover, the ANN architecture has been trained for spring and summer, although
it can also be trained for autumn and winter following the methodology explained
in this section. Hence, as training data set, real data acquired during the operation
of the CDdI-CIESOL-ARFRISOL bioclimatic building, which is described in
Chap. 2 , has been used. More specifically, the real data used as training data set
has been selected to cover the most characteristic profiles of each season, and thus
it comprises three different intervals along the spring and the summer seasons:
from 15 April 2013 until 5 May 2013, from 22 May 2013 until 16 June 2013, and
from 7-31 July 2013 with a sample time of t s = 60 s, and thus, their size is 99,360

data points. Besides, the training data set has been split into the following subsets:

- A training set for obtaining both the ANN parameters using a gradient descent
algorithm. This data set represents the 85% of the global training data set, i.e.
84

456 data points.
- A testing set for avoiding overtraining. This data set is represented by the 15%
of the global training data set which is equivalent to 14,904 data points.

,

Nevertheless, the choice of the input variables is a key factor within the modelling
of temporary data serials. The selection of input variables has been performed as a
function of the prior knowledge of the modelled system. Moreover, the input vari-
ables selected to determine each of the submodels are explained in Sects. 4.2.2.1
to 4.2.2.3 , since they are different as a function of the modelled variable.

•

Training of the ANN . Once that the inputs have been established the training
process is similar for all the submodes. More specifically, it is performed using a
variable-step gradient descent process, namely the MATLAB's implementation of
the Levenberg-Marquardt algorithm (Moré 1978 ). Besides, the trainlm function
is used for at most 150 iterations over the training data set. Furthermore, all the
components of the input vector are normalised by subtracting the mean value and
dividing by the standard deviation for each variable. Afterwards, the goodness of
fit is estimated using the Root-Mean-Square (RMS), see Eq. 4.18 :

N

x

,Ψ) 2

1
N

RMS = ( i ) − ˆ x ( i (4.18)

i = 1

where x ( i ) represents the correct value of the modelled variable and x ˆ ( i ,Ψ) is the

approximation given by the ANN model.
To select an appropriate ANN it should be taken into account that an ANN without
enough nodes may be unable to reproduce the dynamic behaviour of the modelled
variables, and that an ANN with more nodes than needed can cause overtraining,
and thus, degrade the generalisation capabilities. Therefore, the selection of the
network size has been performed by training several ANN with different values of
N h . Finally, the values chosen for each submodel are specified in Sects. 4.2.2.1 -
4.2.2.3 .