面向海上风电设备的短时强对流天气预测模型设计  

Design of short⁃term strong convective weather prediction model for offshore wind power equipment

作　　者：张雪松 王雁冰 郭新毅 ZHANG Xuesong;WANG Yanbing;GUO Xinyi(CGN New Energy Holdings Co.,Ltd.,Beijing 100071,China)

机构地区：[1]中国广核新能源控股有限公司,北京100071

出　　处：《电子设计工程》2025年第9期22-26,共5页Electronic Design Engineering

基　　金：中广核尖峰计划项目(001-GN-A-2021-SN-0239)。

摘　　要：为提升海上风电设备运维安全及短时气象预测精度,文中对深度学习理论和雷达信号处理方法展开研究。利用卷积神经网络(CNN)提取雷达回波图像的空间特征,以长短时记忆网络(LSTM)提取雷达回波信息的时间特征,设计了一个Conv-LSTM神经元处理结构。该结构采用卷积运算替代传统LSTM单元中的矩阵点乘运算,使得二维信息可以在记忆单元中进行映射。此外,为避免高维信息在网络训练时的过拟合现象,引入了一种注意力转移机制,对输入Conv-LSTM网络的信息进行筛选,提升了网络训练效率。基于某型多普勒雷达回波图像进行的仿真验证结果表明,所提算法的POD、FAR、CSI分别达到了0.865、0.199和0.827,较传统的LSTM、CNN网络有明显提升。In order to improve the operation and maintenance safety of offshore wind power equipment and improve the accuracy of short-term weather prediction,deep learning theory and radar signal processing methods are studied in this paper.A Conv-LSTM neural processing structure was designed to extract spatial features of radar echo images using Convolutional Neural Networks(CNN)and temporal features of radar echo information using Long Short-Term Memory(LSTM)Networks.This structure uses convolution operations to replace matrix point multiplication operations in traditional LSTM units,allowing two-dimensional information to be mapped in memory units.In addition,to avoid overfitting of high-dimensional information during network training,an attention transfer mechanism was introduced to filter the information input to the Conv-LSTM network,which improved the efficiency of network training.The simulation verification results based on a certain type of Doppler radar echo image show that the POD,FAR and CSI of the proposed algorithm reach 0.865,0.199 and 0.827,respectively,which is significantly improved compared to traditional LSTM and CNN networks.

关 键 词：LSTM CNN 多普勒雷达 气象预测 海上风电 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]