Computational & Technology Resources
an online resource for computational,
engineering & technology publications
Civil-Comp Proceedings
ISSN 1759-3433
CCP: 78
Edited by: B.H.V. Topping
Paper 51

Artificial Neural Network Modelling of Runoff from Storms in Urban Areas

J. Yang and M. Bruen

Centre for Water Resources Research, Civil Engineering Department, University College Dublin, Ireland

Full Bibliographic Reference for this paper
J. Yang, M. Bruen, "Artificial Neural Network Modelling of Runoff from Storms in Urban Areas", in B.H.V. Topping, (Editor), "Proceedings of the Seventh International Conference on the Application of Artificial Intelligence to Civil and Structural Engineering", Civil-Comp Press, Stirlingshire, UK, Paper 51, 2003. doi:10.4203/ccp.78.51
Keywords: artificial neural network, rainfall-runoff modelling, isolated storm event, discontinuous data, urbanisation, training.

Artificial neural networks (ANN) have been widely used in recent years in hydrology and water resources. As a rainfall-runoff model, ANN has been proven to perform well compared to the conventional time series models, e.g. ARMA models. Although there is little doubt of its usefulness in modelling rainfall-runoff process, the capability of ANN to model discontinuous time series data, such as individual storm event with short time steps, has rarely been discussed. In this paper, the ability of ANN to model discontinuous time series data, i.e., individual, separated storm events, is examined. For the purpose of comparison, the ARMA (p,q) time series models of Box and Jenkins [1] are used as benchmark models. Two ANN modelling tests were carried out on 5-minute data from a small urbanising catchment with a very quick response to rainfall. Different configurations of data inputs, and numbers and arrangement of neurones were tested. The output is the flow in a channel draining the catchment. The urban index, in term of urbanisation rate is used as an extra input variable to the model in order to reflect the land use change due to urbanisation. The following conclusions may be drawn from the study:

  1. ANN performed well and gave consistent results in general. With respect to the time series models of ARMA, ANN provided considerably better performance in terms of R2 efficiency and MSE. The results show that it is possible, using relatively simple ANN structures, to model discrete rainfall-runoff events.

  2. ANN has powerful pattern recognition abilities and is able to capture the dynamics in individual storm events (in both calibration and verification modes). It is encouraging that ANN fits well the observed discharge for both multi-peak events and individual flash floods. It demonstrates that ANN is able to capture individual storm events with short response times if careful training is performed.

  3. The test of "sparse" data confirms that ANN is able to calibrate the model with small sample size. It is important to have sufficient information provided for the pattern recognition in order to generalise the trained model.

  4. The use of an urbanisation index as an extra input variable provided better but not a significant improvement in model performance. Nevertheless, it provides a possible way for incorporating land use change information into the model structure. However, its physical meaning in the context of the network should be further investigated.

Box, G.E.P., Jenkins, G.M., "Time series analysis - forecasting and control, (Revised edition)", Holden-Day Inc. 1976.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £82 +P&P)