Fuzzy neural network and LLE Algorithm for forecasting precipitation in tropical cyclones: comparisons with interpolation method by ECMWF and stepwise regression method
Fuzzy neural network and LLE Algorithm for forecasting precipitation in tropical cyclones: comparisons with interpolation method by ECMWF and stepwise regression method
复制标题
模糊神经网络和 LLE 算法预测热带气旋降水：与 ECMWF 插值法和逐步回归法的比较
DOI：
10.1007/s11069-017-3122-x
复制
发表时间：
2018
期刊：
影响因子：
3.7
通讯作者：
Huang Xiao-yan
中科院分区：
文献类型：
--
作者：
Huang Ying;Jin Long;Zhao Hua-sheng;Huang Xiao-yan
本文以NCEP/NCAR再分析数据的物理量为潜在预测因子，利用模糊神经网络（FNN）模型建立了热带气旋降水预报方案。利用1980-2015年影响广西的172个热带气旋（TC）的TC降水样本进行模式开发。FNN模型输入由潜在的预测因子组成，采用逐步回归方法（SRM）和局部线性嵌入（LLE）算法。LLE算法能够发现隐藏在其非线性高维数据空间中的有意义的低维架构，并分离其潜在因素。在本方案中，采用新建立的FNN-LLE模式，对广西89个站点前一天20:00（北京时间）至当天20:00（北京时间）逐日的TC降水进行预报。利用欧洲中期天气预报中心（ECMWF）的细网格数据，比较了FNN-LLE模型与广泛使用的SRM和插值方法在广西89个站点的TC降水预报效果。采用均方根误差（RMSE）、偏倚和公平威胁评分（ETS）结果来评估预测结果。结果表明，FNN-LLE模型预测TC降水的RMSE值分别为21.94、24.07和25.22，优于ECMWF和SRM插值方法。此外，平均偏差和ETS值接近1.0的FNN-LLE模型的预测效果优于ECMWF和SRM插值方法。
A tropical cyclone (TC) precipitation prediction scheme has been developed based on the physical quantities of the NCEP/NCAR reanalysis data as potential predictors and using fuzzy neural network (FNN) model. TC precipitation samples from 172 tropical cyclones (TCs) affecting Guangxi, China, spanning 1980–2015 are used for model development. The FNN model input is constructed from potential predictors by employing both a stepwise regression method (SRM) and a locally linear embedding (LLE) algorithm. The LLE algorithm is capable of finding meaningful low-dimensional architectures hidden in their nonlinear high-dimensional data space and separating the underlying factors. In this scheme, the newly developed model, which is termed the FNN–LLE model, is used for daily TC precipitation prediction from 20:00 (Beijing Time, or BT) of the previous day to 20:00 BT of the current day at 89 stations covering Guangxi, China. Using identical modeling samples and independent samples, predictions of the FNN–LLE model are compared with the widely used SRM and interpolation method using the fine-mesh data of the European Centre for Medium-Range Weather Forecasts (ECMWF) in terms of the performance of TC rainfall prediction at 89 stations in Guangxi. The root-mean-square error (RMSE), bias, and equitable threat score (ETS) results were employed to assess the predicted outcomes. Results show that the FNN–LLE model is superior to the interpolation method by ECMWF and SRM for TC precipitation prediction with RMSE values of 21.94, 24.07, and 25.22 in FNN–LLE model, interpolation method by ECMWF and SRM, respectively. Moreover, FNN–LLE model having average bias and ETS values close to 1.0 gave better predictions than did the interpolation method by ECMWF and SRM.