Nonlinear Filtering, Parameter Estimation and Decomposition of Large Rainfall-Runoff Models

Abstract

Three topics related to the real time forecasting of river flows are studied. First, the usefulness of nonlinear filtering procedures in connection with a conceptual rainfall-runoff model is investigated. By means of a case study it is determined that only filters which employ future information to correct the past (smoothers) could potentially improve forecasts over the simpler extended Kalman filter. The quality of the predictions is heavily dependent on the nature of the assigned error of the conceptual rainfall-runoff model. The second topic deals with the estimation of the conceptual model error using the maximum likelihood method and consistency conditions on model residuals. The utility of the procedures is tested in practical applications. It is shown the simplified maximum likelihood procedure gives excellent forecasting results independent on the initial conditions, but raises some questions as to the sensitivity of the soil moisture accounting part of the model. - The third topic deals with the forecasting on a basin composed of several interconnected sub-basins. Decomposition procedures are proposed to forecast on sub-basins separately, using upstream flow predictions as inputs to downstream basins. When tested in practice, these methods provide reliable and inexpensive forecasts.