基于多尺度空间相关的气象要素预测

Meteorological Elements Forecasting Based on Multi-scale Spatial Correlation

气象要素预测与人们的日常生活密切相关,为了探索气象要素的空间相关性质,进而实现不同时间尺度的气象要素预测,采用了图信号方法,利用能够简洁地刻画变量间依赖关系的概率图模型,由各个气象观测站点数据构成图模型的节点,建立条件高斯图模型,学习图中边的分布规律,以确定站点间的空间相关关系。根据天气系统的特征尺度,对条件高斯图模型进行改进,引入相似度矩阵,通过交叉验证的方法确定各个预测时间尺度所对应的相似度矩阵参数,建立多尺度空间相关模型,利用改进的二阶有效集方法进行模型估计。将此模型应用于全球3431个国际交换站,分别进行了6小时、24小时、72小时和7天降水量与气温的气象要素预测。实验结果表明,该模型在计算效率和预测准确率方面都有所提升。

Meteorological element prediction is closely related to people’s daily life.In order to explore the spatial correlation properties of meteorological elements and then realize the prediction of meteorological elements at different time scales,the graph signal method is adopted.The probability graph model which can concisely depict the dependence relationship between variables is used.The node of the graph model is composed of the data of each meteorological observation station,and the conditional Gauss graph model is established.To determine the spatial correlation between sites,the distribution of edges in the exercise is studied.According to the characteristic scale of weather system,the conditional Gaussian graph model is improved,and the similarity matrix is introduced.The parameters of the similarity matrix corresponding to each prediction time scale are determined by cross-validation method.A multi-scale spatial correlation model is established,and the model is estimated by using the improved second-order effective set method.The model is applied to 3431 international exchange stations in the world,and the meteorological elements of precipitation and temperature are predicted for 6 hours,24 hours,72 hours and 7 days respectively.The experimental shows that the model has improved both in calculation efficiency and prediction accuracy.