Model Data Selection and Data Pre-processing Approaches

Abstract

Data-based modeling relies on historical data without directly taking account of underlying physical processes in hydrology . So, real-world modeling of hydrological processes commonly requires a complex input structure and very lengthy training data to represent inherent complex dynamic systems. In cases where a large amount of input data is available, and all of which used for modeling, technical issues such as the increase in the computational complexity and lack of memory spaces have been observed. The likelihood of these problems occurring is much greater in the case of hydrological modeling, as these models possess high nonlinearity and a large number of parameters. Therefore, there is a definite need to identify proper techniques which adequately reduce the number of inputs and the required training data length in nonlinear models. Removing redundant inputs from all available input pools and deciding upon the optimum data length to make a reliable prediction are the main purposes of these approaches. This section of the book describes the abilities of novel techniques such as Gamma Test (GT), entropy theory (ET), Principle Component Analysis (PCA), cluster analysis (CA), Akaike’s Information Criterion (AIC ), and Bayesian Information Criterion (BIC ) in model data selection. The novelty of this work is that many of these approaches are used for the first time in hydrological modeling scenarios such as solar radiation estimation, rainfall-runoff modeling , and evapotranspiration modeling . Towards the end of this chapter, conventional data selection procedures such as the Cross-Correlation Approach (CCA), Cross-Validation Approach (CVA), and Data Splitting Approach (DSA) are explained in detail. These traditional approaches were used to check the authenticity of the newly applied methods in the later case study chapters.