Optimization and Variants of Quantile-Based Methods for Bias Corrections of Statistically Downscaled Precipitation Data

Abstract

New optimization and variants of quantile-based methods are developed for bias corrections of monthly and daily general circulation model (GCM)-based statistically downscaled precipitation data. These methods use optimization formulations involving several linear and nonlinear corrections with single and multiple objectives and integrate artificial neural networks (ANNs) with quantile matching (QM) methods. The proposed methods were evaluated at 18 rain gauge sites in Florida using several error and performance measures. Downscaled monthly precipitation data are derived from two statistical downscaling models, including a support vector machine (SVM)-based method developed in this study. Downscaled daily precipitation data from two different climatic zones are also used for the evaluation of bias-correction methods. The methods are assessed based on several performance and error measures, along with their ability to replicate all the moments of the distribution. The selection of the best method among several others for a specific site was found to be dependent on specific performance and error measures adopted for evaluation. The proposed methods not only replicated the observed precipitation data distributions but also minimized the quantitative errors between observed and downscaled precipitation data sets, which could not be accomplished using existing methods. ANN-based methods performed better than QM-based ones in replicating extreme precipitation indices at a daily temporal scale. The multiobjective optimization methods require careful selection of objectives and assignment of weights, with the latter heavily influencing the performance of methods. Variation in performances of methods is observed when methods are calibrated with varying baseline periods with a constant length of test data.