Hydrological Applications of Artificial Neural Networks

Abstract

This paper looks at two example applications of Artificial Neural Networks (ANNs) to hydrology. The first implements a Multi-Layer Perceptron (MLP) to correct flow-rate simulations from the WRIP simulator (Han, 1991) for hourly observations of a single flow-rate and to predict it up to 5-hours in advance. Improvements in accuracy as compared with the WRIP output were clearly demonstrated. The second implements a Multi-Layer Perceptron (MLP) to act as a surrogate for the 3DNet / SIPSON simulator (University of Belgrade, 2010), which uses a conventional fluid-dynamic approach to predict flooding surcharge levels in sewer networks. An MLP-based data-driven model is created to emulate the SIPSON outputs for a 123-manhole sub-network from Keighley, West Yorkshire, UK under a range of rainstorm durations and return periods. A 3-minute sampling period was used. Both a flood level analogue and a classification scheme were successfully implemented. Early results show nowcasting predictive capability for up to 30-minutes ahead.