摘要
针对现有大多数基于循环神经网络的雷暴预测模型普遍存在梯度爆炸、长时依赖关系捕捉能力薄弱以及因无法并行计算导致的效率低下等问题,提出三维时域卷积神经网络(three-dimensional time-domain convolutional neural network,3D-TCN)用于气象雷达雷暴预测。在TCN核心的残差块中引入三维卷积操作,从雷达图像中提取时空特征,实现最长1 h的精准预测。基于数据集对3D-TCN模型的雷暴演变预测性能进行验证,并与时空记忆解耦的空时长短时记忆(spationtemporal long short-term memory,ST-LSTM)网络、记忆增强网络及传统交叉相关算法展开对比。实验结果表明,3D-TCN模型在临界成功指数等评价指标上表现更优,且凭借轻量化结构大幅缩短训练时间。与时空记忆解耦的ST-LSTM网络相比,其临界成功指数平均提升0.037,最高提升幅度接近0.13,训练时间则从4~6 min缩短至1 min以内,验证了所提方法的有效性。
To address the problem of gradient explosion,weak ability to capture long-term dependencies,and low efficiency caused by the inability to perform parallel computing in most existing thunderstorm prediction models based on recurrent neural networks(RNN),a three-dimensional time-domain convolutional neural network(3D-TCN)is proposed for meteorological radar thunderstorm prediction.A 3D convolution operation is introduced into the residual block of TCN core to extract spatiotemporal features from radar images,achieving accurate prediction for up to 1 h.The thunderstorm evolution prediction performance of the 3D-TCN model is validated based on the Hong Kong Observatory 7(HKO-7)dataset,and is compared with the spatiotemporal memory decoupled spatiotemperal long short-term memory(ST-LSTM)network,memory enhancement network,and traditional cross-correlation algorithms.The experimental results show that the 3D-TCN model performs better in evaluation metrics such as critical success index,and significantly reduces training time with its lightweight structure.Compared with the ST-LSTM network decoupled from spatiotemporal memory,its critical success index has increased by an average of 0.037,with the highest increase approaching 0.13.The training time has been shortened from 4 min to 6 min to less than 1 min,which verifies the effectiveness of the proposed method.
作者
陆恩绩
汪玲
朱岱寅
周晔
LU Enji;WANG Ling;ZHU Daiyin;ZHOU Ye(College of Electronic and Information Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China;Key Laboratory of Radar Imaging and Microwave Photonics,Ministry of Education,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China;Leihua Electronic Technologv Research Institute,Aviation Industry Corporation of China,Wuxi 214063,China)
出处
《系统工程与电子技术》
北大核心
2025年第12期4084-4092,共9页
Systems Engineering and Electronics
基金
工信部民机项目资助课题。
关键词
时序卷积网络
气象雷达
雷暴预测
深度学习
temporal convolutional networks(TCN)
weather radar
thunderstorm prediction
deep learning
