NONLINEAR FLOOD FORECASTING BY THE FILTER SEPARATION AR METHOD

At each time step, the runoff time series sampled at equal time intervals are separated through a numerical filter, into the shorter- and longer-period components. Each subsystem is linear and expressible by the ARX method (autoregressive model with exogenous input) or by its conversion form of convolution integral. The effective rainfall components for each subsystem are inversely estimated sequentially from the separated time series of a runoff component, either directly from the ARX expression or solving the simultaneous equations of the convolution integral type, by LP method under the restriction of positive effective rainfalls and minimal errors. Future runoff is predicted with the input of future rainfall for the subsystems either by extrapolating the inversely estimated effective rainfall or by utilizing the given observed rainfall data.