Skip to main content
ARS Home » Pacific West Area » Tucson, Arizona » SWRC » Research » Publications at this Location » Publication #158383
Title: NEURAL-NETWORK-BASED ALGORITHMS OF HYDRAULIC ROUGHNESS FOR OVER LAND FLOW 1545

Author
item LOPEZ-SABATER, C. - UNIVERSITY OF ARIZONA
item RENARD, K. - RETIRED ARS
item LOPES, V. - UNIVERSITY OF ARIZONA

Submitted to: Transactions of the ASAE
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 12/19/2002
Publication Date: 5/20/2003
Citation: Environ. Modelling and Software 18:825-830.

Interpretive Summary: This study presents a series of algorithms developed to estimate hydraulic roughness coefficients for overland flow. The algorithms are combinations of neural networks that use surface configuration parameters and the local flow Reynolds number as inputs, and provide an estimate of the roughness coefficient (Darcy-Weisbach, Manning, or Chezy). Results presented here show that as new neural networks are combined into the stacked algorithm, the estimate errors become gradually smaller. The Final Prediction Error index has been used to identify the optimum network size. Additionally, the dataset used to develop the neural networks, developed from measurements taken at approximately equal Reynolds number intervals, has been found to benefit the algorithms predicting Chezy coefficients. The scarcity of data points in some regions of the output space for the Darcy-Weisbach and Manning models caused a reduction in the predictability of the algorithms for these regions and prevented the use of more complex neural networks. The algorithms have been tested for a wide range of input variables in a detailed sensitivity analysis and have produced reasonable results in all cases.

Technical Abstract: This study presents a series of algorithms developed to estimate hydraulic roughness coefficients for overland flow. The algorithms are combinations of neural networks that use surface configuration parameters and the local flow Reynolds number as inputs, and provide an estimate of the roughness coefficient (Darcy-Weisbach, Manning, or Chezy). Results presented here show that as new neural networks are combined into the stacked algorithm, the estimate errors become gradually smaller. The Final Prediction Error index has been used to identify the optimum network size. Additionally, the dataset used to develop the neural networks, developed from measurements taken at approximately equal Reynolds number intervals, has been found to benefit the algorithms predicting Chezy coefficients. The scarcity of data points in some regions of the output space for the Darcy-Weisbach and Manning models caused a reduction in the predictability of the algorithms for these regions and prevented the use of more complex neural networks. The algorithms have been tested for a wide range of input variables in a detailed sensitivity analysis and have produced reasonable results in all cases.