SPECTRAL ANALYSIS AND FORECASTING OF HYDROLOGIC TIME SERIES

ALI DURGUNOGLU, Purdue University

Abstract

Two important aspects of hydrologic analysis--the spectral analysis and forecasting hydrologic time series--are considered in this study. Five recently developed parametric spectral estimation methods, which are claimed to be without the major drawbacks of the conventional spectrum estimation methods, are analyzed in this study. These spectral estimation methods are proposed by Marple, Kay, Cadzow, Friedlander and Pisarenko. These methods are analyzed first by using synthetic data and then by using real data. The models are selected for these methods by using three objective model selection criteria: Akaike's information, Bayesian, and Hannan-Quinn criteria. Daily rainfall runoff is forecast by using both stochastic and deterministic models. The parameters of the stochastic models (autoregressive moving average -ARMA, threshold autoregressive -TAR, and ARMA models with exogenous inputs -ARMAX) are estimated by a recursive maximum likelihood (RML) method. The deterministic model used in the present study is an improved constrained linear system (CLS) model. The parameters of the CLS model are estimated by using quadratic programming. Friedlander's and Pisarenko's methods are recommended for spectral analysis of hydrologic time series. For forecasting daily rainfall-runoff processes, ARMAX and improved CLS models are found to be superior to the other methods.

Degree

Ph.D.

Subject Area

Hydrology

COinS