Author(s): Ramesh Teegavarapu; Alexis Schauer; Priyank Sharma

Linked Author(s): Ramesh Teegavarapu

Keywords: Model-Tree; Hydroclimatic variables; Regression; Data-driven models

Abstract: Model Tree (MT)-based approaches as emerging data-driven hierarchical methods characterize inter-variable relationships by dividing the input parameter regions into several sub-regions and formulating a multi-variable linear regression model for each sub-region. The MT-based models show an advancement over the classification and regression tree models and many data-driven paradigms. In this study, several MT-based models are evaluated for their ability to forecast multiple hydroclimate variables (viz., temperature, precipitation, and streamflows) at different temporal scales and in two climatic regions. Daily and monthly hydroclimatic variable data from two regions (the U.S. and India) are used for the development of the models. Results from MT-based models are also compared with those from naïve, traditional multiple regression, artificial neural networks, and other data-driven approaches when applied to the prediction of the hydroclimatic variables. A comprehensive evaluation of the models using several error and performance measures is carried out. The efficacy of the MT-based approach for forecasting at different temporal scales, utility for adaptive forecasting applications is evaluated. Improvements in the MT-based estimates considering clustering/stratification of the datasets, post-forecast corrections using domain-specific knowledge, using information about the non-stationary statistical characteristics of time series are evaluated.

DOI: https://doi.org/10.3850/IAHR-39WC2521711920221040

Year: 2022