Rainfall forecasting through artificial neural networks

A knowledge of rainfall is an important component of the information needed for effective flood management. For those catchments with a short response time, forecasting of likely rainfall is necessary for the opportunity to implement relevant action plans. Due to the complexity involved with incorporation of physical parameters in a process model, forecasting rainfall with a process model generally is not possible. Artificial neural networks are an alternative technique. Presented herein is the development of an artificial neural network for forecasting rainfall over the Upper Parramatta River Catchment in Sydney. In addition to discussing the accuracy of predictions obtained, the alternative forms of artificial neural networks (multi-layer feed forward neural network, partial recurrent neural network, and time delay neural network) and the appropriateness of each alternative form for rainfall forecasting will be discussed. Particular attention is given to the order of the lag and the complexity of the networks.