Flow Forecasting Using Artificial Neural Network and a Distributed Hydrological Model, MIKE SHE
Sahoo, Goloka Behari, and Chittaranjan Ray
Book Chaper 10, New Topics in Water Resources Research, Nova Science Publishers, 315–333 pp. http://www.novapublishers.org/catalog/product_info.php?products_id=6191 (2008)
Many drainage basins in Hawaii, especially those on Oahu, are highly urbanized with large areas of impervious surface that produce nearly 100% runoff. Streamflows change by a factor of 60 in only 15 minutes during flashfloods, because these streams are short and steep and respond quickly to intense tropical storms. No matter how small is the scale, many tropical storms on Oahu lead to flash foods and cause life and property damage. Thus, forecasting and analysis of runoff and streamflow in relation to the size and land-use characteristics of individual drainage basins are important for flood and land resources management. This study presents flow estimation of a tropical mountainous stream on Oahu at a 15-minute frequency using a back-propagation neural network (BPNN) and a physically. distributed model MIKE SHE (DHI, 2003). It was found that BPNN was able to predict the streamflow with a correlation coefficient (R) greater than 0.99. It was demonstrated that BPNN was found to be superior to MIKE SHE in terms of predictive performance efficiency: R, mean error, and root mean square error when all information at the gauging stations: stream stage, rainfall, and evapotranspiration was available. MIKE SHE produced continuous and consistent results along the streams (gauged and ungauged locations) with R approximately 0.7. Also, MIKE SHE produced results at multiple sites (i.e. at gauged and ungauged stations) and at multiple scales (e.g., streamflow, groundwater head, surface water level, and soil moisture content). While BPNN is not intended as a substitute for a conceptual model, it can be used as a viable alternative to a physically-distributed model only if streamflows at gauging sites are required. The advantages and limitations of using empirical BPNN and a distributed physical based model for flow forecasting are discussed.