Investigation of Daily Flow Forecasting Models

Abstract

Two types of models for forecasting daily flows from a basin are investigated. These model types are a segmented model whose components are multi-input linear models of the basin between gaging stations, and an unsegmented or fully lumped stochastic basin model of the ARMA (p,q) family. Three approaches to segmented model construction are investigated. The results show that a constrained linear systems estimation method gave models which produced forecasts with a smaller bias than the ordinary least-squares method. The addition of an error model further improved the forecasting performance of the segmented models. A fully lumped stochastic model of the basin is developed and its forecasts are compared with those given by the segmented models. The results show that the segmented models provide better forecasts but exhibited more bias when compared to the unsegmented model of the basin. The results are obtained by using the data from the Green River Basin in Kentucky.