Abstract:

Abstract:Based on regular observation data of marine meteorology,three kinds of numerical forecast products, including Tianjin meso-scale weather research and forecasting （TJ-WRF） model, EC  and its ensemble forecast products from 2015 to 2017, two level sea wind prediction model of BP neural network were established.Firstly, the model was trained with a large number of historical samples to realize the nonlinear mapping between small wind, gale and related factors, respectively. Its results had high prediction accuracy. During the one-year business trial period, the prediction capacity of this model for each wind level and every period of validity of forecast were basically stable. At the same time, the prediction error of the BP neural network was smaller than that of the numerical model before the use of the interpretation technology. The average absolute error of wind speed within 72 hours was about 1.7 m·s-1. The prediction accuracy of the BP neural network for the disaster gale could still be maintained at a high level, and the average absolute error of the wind  with force scale 8 was only 1.77 m·s-1.

Key words: statistical interpretation of NWP products, BP neural network, sea wind foreca

摘要：

基于2015—2017年常规海洋气象观测资料、天津中尺度天气预报模式（TJ-WRF）预报产品、EC数值预报及其集合预报产品，建立了渤海BP神经网络两级海风预报模型，该模型在大量历史样本的拟合训练基础上分别实现小风和大风与相关因子间的非线性映射，其结果有较高的预报准确率。一年的业务试用期间，该模型对各风级、各预报时效的预报能力基本稳定，预报误差较使用该释用技术前数值模式误差有所减小，72 h内风速平均绝对误差为1.72 m·s-1；对灾害大风的预报准确率仍能保持较高水平，8级风风速平均绝对误差仅为1.77 m·s-1。

关键词: 数值预报统计释用, BP神经网络, 两级海风预报模型, 渤海