Abstract:
Artificial neural networks (ANNs) are powerful tools for the modeling and
forecasting of complex engineering systems and have been exploited by researchers to
solve a variety of problems over the last couple of decades. In spite of their proven ability
to provide superior model performance compared to traditional modeling approaches, they
have not become popular among decision makers for operational use. It is probably
because of their perceived black box nature that does not explain or consider the
underlying physical processes involved. This paper presents the results of a study aimed at
a systematic dissection of the massively parallel architectures of trained ANN hydrologic
models to determine if they learn the underlying physical subprocesses during training.
This has been achieved using simple qualitative and quantitative techniques. The data
derived from three contrasting catchments at two different time scales were employed to
develop ANN models and test the methodologies employed for knowledge extraction. The
results obtained in this study indicate that the number of hidden neurons determined
during training for a particular data set correspond to certain subprocesses of the overall
physical process being modeled. It has been found that the time scale of the data employed
has an effect on optimum ANN architecture and knowledge extracted.