Artificial neural network for estimating annual runoff

The first step towards the planning and design of any of water resources project is the estimation of dependable yield which in turn requires the annual runoff volume. The determination of reliable dependable yield requires sufficiently long term flow records. But, most of the river basins in India seldom have a well documented flow data. The immediate remedy to such problem is to develop a regression equation linking the annual runoff volume with other available data such as annual rainfall data, catchment characteristics. Such relationship could be made use of for other hydrometeorologically homogeneous catchments. In this study, an attempt has been made to correlate the annual runoff depth with annual rainfall depth and geomorphological characteristics by using artificial neural network. A pattern mapping algorithm based on gradient descent method is used. This neural network architecture is a feed forward, supervised artificial neural network which captures the generalised non-linear functional relationship between input and output. The size of the network suitable to the available data is decided by a trial and error procedure. The superiority of artificial neural network over the regression equation is demonstrated. Dominant characteristics of catchments affecting the annual runoff are also identified by alternatively using different combinations of the geomorphological characteristics.