基于非参数回归模型的局部线性估计云量预报方法研究

A cloud amount forecasting method based on local linear estimation of non-parametric regression model

为研究云量的分布特点,本文利用历史观测资料,对新义州、定海、隆子3站的总云量和低云量进行了正态性检验,结果显示:总云量和低云量均未达到正态指标,具有一定的随机性.因此,在2004-2007年逐年1月T106L19模式产品和单站地面观测资料的基础上,采用适合被解释对象呈非未知分布的非参数方法——局部线性估计方法,选择合适的窗宽和核函数,创建了上述3站总云量和低云量的短期预报模型,包括不同的长度样本序列.同时,为了比较预报效果,还采用适合被解释对象呈正态分布的参数方法——逐步回归法,建立了相应的预报模型,并利用2003年1月1～31日的逐日T106L19模式产品和3站的云量历史观测资料,对各种预报模型进行了试报和效果的检验,结果表明:在3站的总云量、低云量的月平均准确率和月平均平均绝对误差的检验指标中,非参数局部线性估计的预报精度均高于逐步回归方法;使用短样本序列建立的自适应非参数局部线性估计预报模型与采用长样本序列建立的预报模型相比,效果相当.这意味着,在数值预报产品解释应用的云量预报中,非参数局部线性估计方法可以更合理地考虑其时间分布特征,尤其在缺乏较长时间的历史建模样本时,具有良好的应用前景.

To study the distribution characteristics of cloud amount and by using historical station-observed data of Sinuiju, Dinghai, Lhunze, the normality of total cloud amount and low cloud amount is investigated. The results show that, neither total cloud amount nor low cloud amount satisfies the normality indicators, on the contrary, they have certain randomness. Therefore, based on the T106L19 model products and station-observed data in January from 200,1 to 2007. with the method of non-parametric regression （NPR） local linear estimation suitable to distribution-free, by selecting the appropriate bandwidth and kernel function, the short-term forecasting model for total cloud amount and low cloud amount of the mentioned three stations is established, which includes dataset with different sample length. At the same time, in order to verify the forecasting effects of the NPR forecasting model, another short-term forecasting model is also established with the method of the stepwise regression （SWR） suitable to the normal distribution, and the daily dataset of both TI06L19 model products and station-observed cloud mounts are applied to the above two short-term forecast models. The results indicate that, as to the forecast of total cloud amount and low cloud amount for the mentioned three stations, the precision of NPR model is higher than that of SWR model at the aspects of both monthly averaged forecasting accuracy rate and monthly averaged absolute error, and furthermore, with NPR model, the forecasting effects are better with the short period of sample compared with the long period of sample. That means, in the process of interpreting numerical forecasting products, distribution characteristics of sample time series should be considered more rationally for using NPR model, especially in the absence of longer period historical samples. NPR model has a good application prospect.