Partitioned state and parameter estimation for real-time flood forecasting

A new algorithm is developed for simultaneous state-parameter estimation in real-time flood-forecasting applications. Dubbed the partitioned state-parameter (PSP) algorithm, is it unusual in the way that the parameter filter is formulated explicitly in terms of the identifiable parameters in the transition and input coefficient matrices. By virtue of its parallel filter structure the algorithm is very fast, yet it has been designed so that essential error interactions between the forecasting and parameter filters are preserved. Furthermore, PSP is structured so that input coefficients are only updated when the corresponding inputs are actually applied. This feature is useful for systems subject to sporadic inputs. The algorithm is tested with real and synthesized daily rainfall-runoff data from the Hillsborough River in Florida. PSP is found to produce good forecasts and parameter estimates and is much faster than the extended Kalman filter.