Feedforward artificial neural networks (ANNs) that are trained with the bac
k-propagation algorithm are a useful tool for modelling environmental syste
ms. They have already been successfully used to model salinity, nutrient co
ncentrations, air pollution, and algal growth. These successes, coupled wit
h their suitability for modelling complex systems, have resulted in an incr
ease in their popularity and their application in an ever increasing number
of areas. They are generally treated as black box models that are able to
capture underlying relationships when presented with input and output data.
In many instances, little consideration is given to potential input data a
nd the internal workings of ANNs. This can result in inferior model perform
ance and an inability to accurately compare the performance of different AN
N models. Back-propagation networks employ a modelling philosophy that is s
imilar to that of statistical methods in the sense that unknown model param
eters (i.e., connection weights) are adjusted in order to obtain the best m
atch between a historical set of model inputs and corresponding outputs. Co
nsequently, the principles that are considered good practice in the develop
ment of statistical models should be considered. In this paper, a systemati
c approach to the development of ANN based forecasting models is presented,
which is intended to act as a guide for potential and current users of fee
dforward ANNs that are trained with the back-propagation algorithm. Issues
that need to be considered in the model development phase are discussed and
ways of addressing them presented. The major areas covered include data tr
ansformation, the determination of appropriate model inputs, the determinat
ion of an appropriate network geometry, the optimisation of connection weig
hts, and validation of model performance. (C) 2001 Elsevier Science Ltd. Al
l rights reserved.