Pauline Kneale
Stan Openshaw
Adrian McDonald
Steve Carver
1st May 1996 - 31st January 1998
MAFF Award
The research provides an empirical investigation of the potential for using neural network based real time flood forecasting models both as independent models and by comparison with more conventional modelling methods.
The empirical analysis involves comparing neural net models of both the entire flood forecasting process as well as hybrids that combine parts of both technologies. An assessment is made of their data needs, real-time forecasting performance, handling of missing and uncertain (noisy and fuzzy) data, computational speed and efficiency, ability to handle sudden and unexpected changes as well as their economics of running, training, and set-up costs. Answers to these questions are based on a series of carefully selected and controlled empirical experiments using historical data series for a number of river gauging stations on the River Ouse system, Yorkshire and weather stations within radar and other spatial data sources in northern England. The results highlight the principal benefits and limitations of using different types of artificial neural networks. They provide an outline system specification for the technology to be developed further from pilot-prototype into live-operational environments.
For further details see the project webpage. A commercial, windows based, version of the flood predictor is available. Please contact Pauline Kneale.