基于主从时空预测网络的闪电临近预警方法
首发时间:2021-04-08
摘要:闪电作为最常见的自然灾害之一对人的生命以及工业基础设施有重大威胁,精细时空分辨率下的闪电临近预警对这种灾害的预防具有重要意义。然而,由于雷暴发展的不稳定性,该任务在气象预报领域仍具有重大挑战。为此,本文提出了一个主从时空预测网络模型(Master-Slave Spatiotemporal Prediction Network,MSTNet)实现闪电的临近预警。模型的输入包括两种时空数据:历史闪电观测与雷达回波。MSTNet利用主从时空预测网络分别对两种数据特有的时空规律进行挖掘:主预测网络用于建模历史闪电与未来闪电之间的强时序关系;从预测网络则负责推演发展相对稳定的雷达回波,为预测未来闪电提供间接支撑。主从预测网络提取得到的特征通过一个融合模块实现信息互补,基于融合特征完成对闪电有无以及强度的临近预警。在我国华南地区实际采集的数据集随论文一同发布)上进行了实验,结果表明MSTNet相比于其他先进的时空序列天气预报模型具有明显优势。
关键词: 时空序列预测 闪电临近预警 时空编码网络 卷积时空网络 深度学习
For information in English, please click here
A Master-Slave Spatiotemporal Prediction Network for Lighting Nowcasting
Abstract:Lightning is one of common natural disasters, posing major threats to human life and industrial infrastructure. Lightning nowcasting under fine spatiotemporal resolution is of great significance for relieving the losses caused by this disaster. However, the nowcasting task is still facing serious challenges due to the instability of thunderstorm development. In this paper, we propose a Master-slave Spatio-Temporal prediction Network (MSTNet) for lightning nowcasting. The model inputs consist of two spatiotemporal sequences: lightning observation and radar reflectivity in recent period. MSTNet employs a master encoder and a slaver encoder to mine the complementary information underling the two kinds of data, respectively. The master encoder aims at modeling the strong temporal dependency between hisA Master-Slave Spatiotemporal Prediction Network for Lighting Nowcastingtorical lightning and future lightning; the slaver encoder intends to simulate the development of radar echoes, providing potential knowledge for lighting nowcasting. The features extracted by the two encoders are then integrated via a fusion module, based on which the final nowcasting results are generated. Experiments are conducted on a real-world dataset (released with our paper) collected in South China. The experimental results demonstrate that MSTNet achieves state-of-the-art performance compared with several advanced baselines utilized for spatiotemporal series weather nowcasting.
Keywords: spatiotemporal series prediction lightning nowcasting spatiotemporal encoder network convolutional spatiotemporal network deep learning
基金:
引用
No.****
动态公开评议
共计0人参与
勘误表
基于主从时空预测网络的闪电临近预警方法
评论
全部评论0/1000