Rainfall-runoff modelling using neural networks and genetic programming (1997)

Funding body: ERASMUS scheme (European Community)

This project presents an application of Neural Networks (NNs) to rainfall-runoff modelling. Applications of the neural network technique in this domain of hydrology have so far provided accurate results for small storm events on theoretical catchments (Minns & Hall, 1995). The aim of the research presented in this report was to investigate the application of NNs, as 'black-box' models of rainfall-runoff processes, on real catchments. The NN approach is tested and compared to optimised conceptual hydrological models applied to a catchment over a period of several years. At the same time, all tests and experiments were done in parallel with a Genetic Programming technique (Cousin, 1997).

Thus, the performance of both data-driven methods could be compared to the model-driven approach (conceptual models). This report demonstrates how a NN and GP can be set up to obtain the best results given the necessary input data. The study revealed that the choice and preparation of calibration data sets are more important than the fine-tuning of the NN (choice of optimal parameters). Both GP and NN had similar behaviours and the final results were quite close to the model-driven approach results in terms of correlation and possible evaluation parameters.

References

  • Jacq, F. and D.A. Savic, (1997), Rainfall-Runoff Modelling Using Neural Networks, Centre For Systems And Control Engineering, Report No. 97/02, School of Engineering, University of Exeter, Exeter, United Kingdom, p.66.
  • Cousin, N. and D.A. Savic, (1997), A Rainfall-Runoff Model Using Genetic Programming, Centre For Systems And Control Engineering, Report No. 97/03, School of Engineering, University of Exeter, Exeter, United Kingdom, p.70.

Back to Artificial Intelligence research and applications

Google+