Combining rainfall-runoff model outputs for improving ensemble streamflow prediction

This study reviewed various combining methods that have been commonly used in economic forecasting, and examined their applicability in hydrologic forecasting. The following combining methods were investigated: The simple average, constant coefficient regression, switching regression, sum of squared error, and artificial neural network combining methods. Each method combines ensemble streamflow prediction (ESP) scenarios of the existing rainfall-runoff model, TANK, those of the new rainfall-runoff model that has been developed using an ensemble neural network for forecasting the monthly inflow to the Daecheong multipurpose dam in Korea. In addition to the combining, the ESP scenarios were adjusted using correction methods, such as optimal linear and artificial neural network correction methods. Among the tested combining methods, sum of squared error (SSE), a combining method using time-varying weights, performed best with respect to the root mean square error. When SSE was coupled with optimal linear correction (OLC), denoted SSE/OLC, its bias became sufficiently close to zero. SSE/OLC also considerably improved the probabilistic forecasting accuracy of the existing ESP system.