Civil Engineering Reference
In-Depth Information
Simulation
software (or
experiments)
Artificial Neural
Network to be trained
and validated
Database
Non-
dominated
solution
Artificial
Neural Network
(validated)
Genetic
Algorithm
Evaluation of individuals
Fig. 8
GAINN framework (Magnier and Haghighat
2010
)
rapidity of evaluation provided by ANN as well as the optimization power of the
GA. The procedure is to first use an ANN to approximate the system being studied,
and then use this ANN within the GA as the objective function. The outcome is a
drastic reduction in the simulation time, while keeping an acceptable quality and
reliability in the solution process. The complete workflow of GAINN is illustrated
in Fig.
8
and is divided into three steps. First, a base software or experimental
setup is used to generate a database of cases. Once the database is created, it can be
used to train and validate the ANN. The ANN is then integrated into the GA as the
objective function, so the GA can run with almost instantaneous evaluation of
individuals. The GA optimization finally provides the non-dominated solution set
(Magnier and Haghighat
2010
).
GAINN was first used in building engineering for the optimization of chillers
control (Chow et al.
2002
). This study introduced the methodology to the building
field and proved its efficiency in terms of accuracy and reduction in the total
optimization time. Later, GAINN has been successfully applied in other studies,
such as Zhou (
2007
), combined with computational fluids dynamics, and Conraud
(
2008
), combined with ESP-r.
Recently, this approach was used by Magnier et al. (
2010
) using a simulation-
based ANN to characterize building behavior, and then the ANN model was
combined with a multi-objective GA to optimize thermal comfort and energy
consumption in a residential building.
According to the previous studies, the GAINN methodology can be very effi-
cient for building application. Due to the ANN evaluation inside the GA, a sig-
nificant amount of time can be saved, while keeping the optimization reliable. One
main limitation of GAINN is that the optimization results rely on the ANN
accuracy. If the ANN is not 100 % accurate, results could be affected and optimal
solutions could be missed.
Another major drawback regarding how GAINN methodology has been applied
so far is the handling of multiple objectives. In the great majority of previous
