EOF迭代模型的月动力延伸预报产品释用方法

Application Method of Monthly Dynamics Extended-range Forecast Based on EOF Iteration Scheme

以动力延伸预报产品为基础,应用EOF迭代模型对河北省2005—2007年6—8月逐月降水量进行预测试验。经过详细分析不同相关区因子、截取特征向量的项数、EOF迭代次数和迭代误差对预测结果的影响,提出在建立模型时应该尽可能选取与预测对象关系密切的相关区因子。随着迭代次数增加,迭代误差逐渐减小,但预测精度与迭代误差没有必然联系。试验结果表明:迭代次数不宜过多,甚至不需要迭代,当仅取前一项特征向量,不进行迭代时,月降水量预测值和实况值的符号一致率为0.63,PS评分为0.74。

Based on EOF iteration scheme model with dynamic extended-range forecast products, experiment of precipitation prediction is done for 121 station of Hebei Province from June to August in 2005-2007. It is necessary to use 500 hPa correlation area in the EOF iteration scheme model in order to improve the prediction accuracy. Experiment in different correlation areas indicates that the model performs best when the critical correlation coefficient is 0.45, correlation area is of 90 grids. The average sign coincidence rate between predicted precipitation and observed precipitation is 0.54. When critical correlation coefficient is 0.45, grid number is 90 and iteration error is 0.05, let truncation order of eigenvector equal 1,2, …, 11 and do experiment separately. The first two eigenvectors account for about 50% of total variance explained. When truncation order equal 1 and 2, average sign coincidence rate are 0.60 and 0.58. When truncation order is greater than 2, average sign coincidence rate decreases as truncation order increases slightly until stability in 0. 54-0. 52. When critical correlation coefficient is 0.45, grid number is 90 and truncation order of eigenvector is 4, let iteration number equal 0,1,…, 19 do the prediction experiment separately . The experiment shows that iteration error decreases as iteration number increases, but the decrease rate is very different. As accuracy reaches 0.05, the fastest experiment requires only one times of iterative operator and the slowest one requires 15 times. The predicted result affected by iteration error is slight. To acquire good result, many times of iteration is unnecessary, usually no more than three times. When truncating the first eigen- vector without iteration, the average sign coincidence rate of monthly precipitation prediction can reach 0.63 and operational scores PS can reach 0.74.