Rainfall forecastingia important for many catchment management applications, in
particular for flood warning systems. The variability of rainfall in spsce and time, howeve r, renders
quantitative forecasting of rainfall extremely difIicult. The depth of rainfall and its diiribution in
the temporal and spatial dimensions depends on many variables, such ss pressure, temperature, and
wind speed and direction. Due to the complexity of the atmospheric proceesea by which rainikll is
generated and the lsck of available data on the necessary temporal and spatial scales, it is not fessible
generally to forecest rainfall using a physically based process model. Recent developments in artificial
intellllence and, in particular, these techniques aimed at pattern recognition, however, provide en
altemstive approach Tar developing of a rainfall forecasting model. Artlflclal neural networks (ANNE),
which perform a nonlinear mapping between inputs and outputs, are one such technique. Preeented
in thii paper m the reeults of a study investigst~g the application of ANNE to forecast the spatial
distribution of rainfall for an urban catchment. Three alternative types of ANNE, namely multilayer
feedforward neural networks, partial recurrent neural networks, and time delay neural networks, were
identified, developed and, ss presented in this paper, found to provide reasonable predictions of the
rainfall depth one time-step iu advance. The data requirements. for and the accuracy obtainable
from these three alternative types of ANNs are discussed.