基于分位数回归方法的统计降尺度降水预报研究

作者：Reza Tareghian, Peter F. Rasmussen

刊物：《水文学杂志》，2013年4月，487卷，122-135页

关键词：降尺度统计；降水量；分位数回归方法；变量选择

摘要：统计降尺度降水预报研究是许多气候变化研究的重要组成部分。基于回归模型的统计降尺度方法中的一个重要步骤是依据条件分布进行抽样以保证实测数据系列方差的稳定性。本研究中，提出了一种降水降尺度的技术。该技术采用分位数回归方法来确定给定日的降雨条件分布，而不是采用传统的线性回归模型，从而减少了标准线性回归模型中的部分假设约束，例如回归误差呈正态分布且与原分布具有相同误差的假设。同时，分位数回归方法在选择预测变量时具有更大的灵活性，使得不同的预测因子集合可被利用于条件分布的不同地区。在选择预测变量时，一般采用利用适用于分位数回归的贝叶斯方法。研究结果表明，该方法在预测夏季降水量时，比传统回归模型具有更大的优势，但是在预测冬季降水量时两种方法没有很大的区别。

Statistical downscaling of precipitation using quantile regression

Authors: Reza Tareghian, Peter F. Rasmussen

Journal: Journal of Hydrology, volume 487, 22 April 2013, Pages 122–135

Key words: Statistical downscaling; Precipitation; Quantile regression; Variable selection

Abstract：Statistical downscaling of precipitation is required as part of many climate change studies. Statistical downscaling based on regression models requires one to sample from the conditional distribution to preserve the variance of observed precipitation. In this paper, we present a new technique for downscaling precipitation. The proposed method employs quantile regression rather than traditional linear regression models to determine the conditional distribution for a given day. This eliminates the need for some of the assumptions required in standard linear regression, including the assumption of normally-distributed errors with constant variance. The quantile regression model also allows considerable flexibility in selecting predictor variables in that different subsets of predictors can be used for different parts of the conditional distribution. A Bayesian method adapted to quantile regression is used to select predictor variables. The method is illustrated through an application to five weather stations in Canada. It is found that the proposed method has distinct advantages over the conventional regression model for predicting summer precipitation, while for winter precipitation there is not much difference between the two methods.

资料来源：http://www.sciencedirect.com/science/article/pii/S0022169413001522