An Adaptive Algrerithm for Analyzing Short-Term Structural and Parameter Changes in Hydrologic Predict in Models 'Source Title' Modeling Hydrologic Processes, Proceedings Of 3rd International Hydrology Symposium
Bibliographic details
Summary
An adaptive unbiased minimum variance recursive prediction algorithm, based upon the state space description of hydrologic systems, is discussed. Discrete linear systems with white Gaussian disturbances are considered. The algorithm allows for short term structural and parameter changes due to random environmental effects. A prediction model is set up from a representation of the rainfall/runoff processes with the unknown parameters modeled by a random walk. The predictions are obtained by the use of linear Kalman filters where the unknown noise covariance matrices are also adaptively estimated. In the second part of the paper a dynamiic Bayesian model selection technique is discussed. It is shown that the essentially nonlinear problem is decomposable into two parts, a linear nonadaptive part consisting of a set of Kalman filters assigned to each priori chosen model and a Bayesian nonlinear part which incorporates the learning nature of the algorithm. The procedure makes the dynamic model selection/structural adaptivity easy to perform by analyzing the innovation sequences assigned to each postulated model. The behavior of the adaptive prediction algorithm is illustrated by real world example taken from rainfall/runoff flood forecasting.