Keywords

surrogate model; water budget; estimation; regional watershed; artificial neural network

Location

Session H2: Water Resources Management and Planning - Modeling and Software for Improving Decisions and Engaging Stakeholders

Start Date

18-6-2014 10:40 AM

End Date

18-6-2014 12:00 PM

Abstract

We developed a method based on surrogate modelling to estimate the water flow in both surface and subsurface sections of a basin watershed. The real case of a regional watershed near Tokyo with a catchment area of 100km2 was targeted with observations on a period of about 16 years. Our purpose was to propose a tool of water resources management that can be operated daily by the person in charge of a watershed without expert knowledge of the numerical modelling process. Replacement of complex physical models by surrogate models like artificial neural networks (ANN) has been proved useful in several applications in hydrology. We think that it is a primary choice when the computation time and the complexity of the physical model become prohibitive for a usage by non-specialists on a standard computer. Here, five surrogate models were tested with a focus on ANN. A feed forward neural network and a radial basis neural network were trained with the Levenberg- Marquardt algorithm to emulate a general purpose fluid flow simulator with fully-coupled surface/subsurface capability used to simulate water flow volume that cannot be observed directly. The comparison with others surrogate modelling techniques has shown the relative superiority of ANN in this application. One configuration of input datasets was made by selecting meteorological (e.g. temperature, rainfall) and hydrological (e.g. streamflow, groundwater level) time series. A data selection study showed the effectiveness of considering river flow rate and rainfall over the groundwater level as input data. Regarding the output variables, six water budget components were selected to fully characterize the water balance of the watershed. The performances were evaluated by mean of the coefficient of determination (R2). Estimation of the unobserved water balance component provides indicators of the watershed condition and guides the sustainable management of the watershed.

Share

COinS
 
Jun 18th, 10:40 AM Jun 18th, 12:00 PM

Using surrogate modelling for fast estimation of water budget component in a regional watershed

Session H2: Water Resources Management and Planning - Modeling and Software for Improving Decisions and Engaging Stakeholders

We developed a method based on surrogate modelling to estimate the water flow in both surface and subsurface sections of a basin watershed. The real case of a regional watershed near Tokyo with a catchment area of 100km2 was targeted with observations on a period of about 16 years. Our purpose was to propose a tool of water resources management that can be operated daily by the person in charge of a watershed without expert knowledge of the numerical modelling process. Replacement of complex physical models by surrogate models like artificial neural networks (ANN) has been proved useful in several applications in hydrology. We think that it is a primary choice when the computation time and the complexity of the physical model become prohibitive for a usage by non-specialists on a standard computer. Here, five surrogate models were tested with a focus on ANN. A feed forward neural network and a radial basis neural network were trained with the Levenberg- Marquardt algorithm to emulate a general purpose fluid flow simulator with fully-coupled surface/subsurface capability used to simulate water flow volume that cannot be observed directly. The comparison with others surrogate modelling techniques has shown the relative superiority of ANN in this application. One configuration of input datasets was made by selecting meteorological (e.g. temperature, rainfall) and hydrological (e.g. streamflow, groundwater level) time series. A data selection study showed the effectiveness of considering river flow rate and rainfall over the groundwater level as input data. Regarding the output variables, six water budget components were selected to fully characterize the water balance of the watershed. The performances were evaluated by mean of the coefficient of determination (R2). Estimation of the unobserved water balance component provides indicators of the watershed condition and guides the sustainable management of the watershed.