Civil Engineering Reference
In-Depth Information
Fig. 3.1
Typical network scheme for a neural network with two hidden layers
3.3.1.2 Neural Network Approach
ANNs are universal approximators (Cybenko
1989
). The most common type is a
static Feed-Forward (FF) configuration that allows to approximate any nonlinear
staticmapping between input and output variables provided that certain conditions are
met. These conditions, regarding the number of nodes of theANNand the smoothness
of the mapping, are often neglected, relying on a final validating step. This is due to
the fact that these important factors are, inmost cases unknown, leading to an iterative
design procedure. Thus, ANN can be considered a black-box model where the model
inputs are the number of neurons in the input layer, the model parameters are the
number of neurons and the values of interconnection weights, which do not have
any physical meaning, in the hidden layers and, lastly, the outputs are the number of
neurons in the output layer. A typical ANNs scheme with two hidden layers can be
observed in Fig.
3.1
. The flexibility of ANN comes at a price, not only the number
of nodes but also the selection of the weights must be decided by trial and error. In
most applications, the iterative design procedure can be tackled using large amounts
of data that are split into several sets, some of which are used for training/design and
others to validate the solution.
In Atthajariyakul and Leephakpreeda (
2005
) an FF ANN is used to provide an
estimation of Fanger's PMV model. The used variables are air temperature, air wet
bulb temperature, globe temperature, air velocity, clothing insulation and human
activity. In this way, the authors avoid the need for hygrometers, which are com-
plex and costly and provide a means for low-cost real-time control. The training
procedure is based on choosing 2
10
5
data points covering the six-dimensional
.
3
×
