| 基于生成对抗网络GAN的人工智能临近预报方法研究 |
| |
| 引用本文： | 陈元昭,林良勋,王蕊,兰红平,叶允明,陈训来.基于生成对抗网络GAN的人工智能临近预报方法研究[J].大气科学学报,2019,42(2):311-320. |
| |
| 作者姓名： | 陈元昭 林良勋 王蕊 兰红平 叶允明 陈训来 |
| |
| 作者单位： | 深圳市气象局;广东省气象台;哈尔滨工业大学(深圳) |
| |
| 基金项目： | 中国气象局预报员专项（CMAYBY2017-052；CMAYBY2019-081）；深圳市科技创新项目（JCYJ20190422090117011）；广东省气象局科技创新项目（GRMC-2016-04；GRMC2018Z06） |
| |
| 摘 要： | 研究设计了基于生成对抗网络(Generative Adversarial Networks,GAN)的人工智能临近预报方法,并进行了业务试验。该方法利用广东12部S波段天气雷达2015—2017年海量雷达拼图资料进行人工智能学习来做临近预报。GAN方法从一系列雷达观测资料中,运用卷积法提取回波图像信息建立预报模型,并通过损失函数训练模型,得到基于人工智能技术的临近预报。对2018年发生在广东地区的4个天气过程的外推预报试验表明,GAN方法对对流天气过程的回波位置、形状及强度的临近预报多数情况下与实况基本一致,具有较好的预报效果。但是该方法预报的回波范围偏大,对层状云降水的预报效果较差。对西风带系统引起的降水,西南季风降水,东风系统引起的降水以及台风降水的18个个例1 h预报的3个级别的回波强度检验发现,GAN方法对中等强度回波的预报较好,但对强回波的预报效果仍有待提高。
|
| 关 键 词： | 人工智能 生成对抗网络 雷达回波 临近预报 |
| 收稿时间： | 2019/1/17 0:00:00 |
| 修稿时间： | 2019/3/26 0:00:00 |
A study on the artificial intelligence nowcasting based on generative adversarial networks |
| |
| CHEN Yuanzhao,LIN Liangxun,WANG Rui,LANG Hongping,YE Yunming and CHEN Xunlai.A study on the artificial intelligence nowcasting based on generative adversarial networks[J].大气科学学报,2019,42(2):311-320. |
| |
| Authors: | CHEN Yuanzhao LIN Liangxun WANG Rui LANG Hongping YE Yunming and CHEN Xunlai |
| |
| Institution: | Shenzhen Meteorological Bureau of Guangdong, Shenzhen 518040, China,Guangdong Meteorological Observatory, Guangzhou 510080, China,Shenzhen Meteorological Bureau of Guangdong, Shenzhen 518040, China,Shenzhen Meteorological Bureau of Guangdong, Shenzhen 518040, China,Harbin Institute of technology(Shenzhen), Shenzhen 518040, China and Shenzhen Meteorological Bureau of Guangdong, Shenzhen 518040, China |
| |
| Abstract: | Artificial intelligence nowcasting based on generative adversarial networks (GAN) has been conducted by using abundant radar echo images from 12 S-band Doppler radars in Guangdong province during the period from 2015 to 2017.Radar echo images were convoluted for 5 times in order to build the initial forecasting model.Afterwards,several confrontation trainings took place between the model images and real radar echo images,resulting in the loss function.The model was optimized constantly.Given that the model images were similar to the real radar echo images,the outputs of optimum model would be used for nowcasting.The experiments of four precipitation events in Guangdong province during 2018 suggested that the 60 min forecasted position,shape and intensity of radar echo in convective systems by GAN mostly coincide with the observations.However,the forecasted area of strong radar echo is larger than that of the observed radar echo.Furthermore,the GAN method could not forecast the precipitation caused by stratus clouds well.The GAN method could forecast moderate radar echoes quite well,while its forecast capability for strong radar echoes needs to be improved. |
| |
| Keywords: | artificial intelligence generative adversarial networks radar echo nowcasting |
| 本文献已被 CNKI 等数据库收录！ |
| 点击此处可从《大气科学学报》浏览原始摘要信息 |
| 点击此处可从《大气科学学报》下载免费的PDF全文 |
|
