基于LGU-Net的数值天气预报融合订正方法

吴鲲鹏; 袁建富; 赵宇恒; 王璐

doi:10.16032/j.issn.1004-4965.2025.072

基于LGU-Net的数值天气预报融合订正方法

doi: 10.16032/j.issn.1004-4965.2025.072

长光卫星技术股份有限公司, 吉林长春 130000

详细信息

通讯作者:
吴鲲鹏，男，吉林省人，硕士，主要从事深度学习、气象预测等领域的研究。E-mail: 451926128@qq.com

中图分类号: P456.7
计量
- 文章访问数: 11
- HTML全文浏览量: 8
- PDF下载量: 1
- 被引次数: 0
出版历程
- 收稿日期: 2024-07-10
- 修回日期: 2025-05-11
- 网络出版日期: 2026-01-04
- 刊出日期: 2025-12-20

Fusion Correction Method for Numerical Weather Forecast Based on LGU-Net

Chang Guang Satellite Technology Co., Ltd., Changchun 130000, China

摘要

摘要: 数值天气预报是现代天气预报的主流技术方法，近年来一直在向精细化方向发展，但预报误差仍无法完全避免。本研究提出了基于LSTM-GAM-UNet（LGU-Net）的数值预报偏差订正模型，该模型在CU-Net的基础上引入了长短期记忆网络（Long Short-Term Memory, LSTM）结构和全局注意力机制（Global Attention Mechanism, GAM），进一步融合多种气象要素、“吉林一号”卫星获取的地形特征以及卫星云图，构建多要素融合订正模型，并专门针对气象预报领域进行了优化设计。在中国东北地区进行实验，对美国国家环境预报中心的全球预报系统（Global Forecast System, GFS）数值预报模式中2 m温度(T2)、2 m露点温度(D2)、10 m的风速（U10、V10）和降水量进行订正，并进行了不同模型的偏差订正实验和对比分析。通过与GFS数值预报模式原始预报结果、模式距离积分订正预报法（Anomaly Numerical-correction with Observation, ANO）订正结果以及CU-Net方法订正结果进行对比，结果表明LGU-Net模型能有效改进数值预报订正效果。此外，云图数据的加入对降水量订正有着明显的正向增益效果，其均方根误差（RMSE）和平均绝对误差（MAE）相对于GFS分别提升了80.76%和76.04%。本研究为高精度气象要素预报提供了新的技术支持。
- 数值天气预报 /
- 深度学习 /
- LGU-Net /
- 偏差订正
Abstract: Numerical weather forecast is the mainstream technique in modern weather forecasting, which has been developing towards higher resolution in recent years, while forecast errors remain unavoidable. This paper proposes a numerical forecast bias correction model, termed LSTM-GAM-UNet (LGU-Net), which introduces the Long Short-Term Memory (LSTM) structure and Global Attention Mechanism (GAM) based on the CU-Net model. The model further integrates various meteorological elements, terrain features derived from"Jilin-1"satellite data, and satellite cloud images, thereby constructing a multielement fusion correction model. This model is also specifically optimized for meteorological forecasting. An experiment was conducted in northeastern China to correct the biases of the Global Forecast System (GFS) for the 2 m temperature (T2), 2 m dew point temperature (D2), 10 m wind components (U10, V10), and precipitation. Different models were compared and analyzed for bias correction experiments. By comparing with the original GFS forecasts, as well as the correction results from Anomaly Numericalcorrection with Observation (ANO) method and the CU-Net method, it was shown that the LGU-Net model effectively improved the bias correction performance. In addition, the addition of cloud imagery data has a significant positive impact on precipitation correction, with an 80.76% and 76.04% improvement in RMSE and MAE compared to GFS, respectively. This paper provides new technical support for highprecision meteorological element forecasts.
- numerical weather prediction /
- deep learning /
- LGU-Net /
- bias correction

HTML全文

图 1 实验区位置

下载: 全尺寸图片幻灯片

图 2 LGU-Net结构

下载: 全尺寸图片幻灯片

图 3 LSTM模型结构

下载: 全尺寸图片幻灯片

图 4 GAM注意力机制

下载: 全尺寸图片幻灯片

图 5 偏差订正流程图

下载: 全尺寸图片幻灯片

图 6 T2订正误差分布图

下载: 全尺寸图片幻灯片

图 7 降水订正误差分布图

下载: 全尺寸图片幻灯片

表 1 T2订正结果误差评价

评价指标	ANO	CU-Net	LGU-Net(单)	LGU-Net（多）	GFS
RMSE/℃	3.658	2.423	1.832	1.011	4.738
MAE/℃	2.254	1.539	1.214	0.823	3.124

下载: 导出CSV

表 2 D2订正结果误差评价

评价指标	ANO	CU-Net	LGU-Net(单)	LGU-Net(多)	GFS
RMSE/m	3.435	2.950	1.965	1.124	5.245
MAE/m	2.171	1.586	1.277	0.923	3.715

下载: 导出CSV

表 3 U10订正结果误差评价

阈值	评价指标	ANO	CU-Net	LGU-Net(单)	LGU-Net(多)	GFS
/	RMSE/(m·s^-1)	2.658	1.744	1.436	0.850	3.412
/	MAE/(m·s^-1)	1.957	1.330	1.023	0.650	1.825
0.2 m·s^-1	POD/(m·s^-1)	0.72	0.77	0.79	0.88	0.67
	FAR/(m·s^-1)	0.70	0.65	0.58	0.51	0.72
	CSI/(m·s^-1)	0.36	0.39	0.44	0.48	0.34
10 m·s^-1	POD/(m·s^-1)	0.19	0.32	0.35	0.40	0.14
	FAR/(m·s^-1)	0.68	0.56	0.46	0.34	0.78
	CSI/(m·s^-1)	0.22	0.25	0.29	0.30	0.17

下载: 导出CSV

表 4 V10订正结果误差评价

阈值	评价指标	ANO	CU-Net	LGU-Net（单）	LGU-Net(多)	GFS
/	RMSE/(m·s^-1)	2.374	1.678	1.396	0.894	3.737
/	MAE/(m·s^-1)	1.850	1.265	0.987	0.671	1.730
0.2 m·s^-1	POD/(m·s^-1)	0.73	0.76	0.81	0.89	0.69
	FAR/(m·s^-1)	0.65	0.61	0.58	0.49	0.70
	CSI/(m·s^-1)	0.34	0.39	0.47	0.51	0.32
10 m·s^-1	POD/(m·s^-1)	0.20	0.34	0.33	0.41	0.15
	FAR/(m·s^-1)	0.63	0.54	0.46	0.33	0.77
	CSI/(m·s^-1)	0.23	0.27	0.31	0.37	0.19

下载: 导出CSV

表 5 降水量订正结果误差评价

阈值	评价指标	ANO	CU-Net	LGU-Net（无云图）	LGU-Net（云图）	GFS
/	RMSE/(mm·h^-¹)	3.374	3.005	1.367	0.912	4.741
/	MAE/(mm·h^-¹)	2.150	2.074	1.170	0.633	2.642
0.1 mm·h^-¹	POD/(mm·h^-¹)	0.79	0.85	0.89	0.91	0.70
	FAR/(mm·h^-¹)	0.68	0.64	0.59	0.48	0.72
	CSI/(mm·h^-¹)	0.28	0.31	0.39	0.50	0.24
7.0 mm·h^-¹	POD/(mm·h^-¹)	0.19	0.25	0.30	0.42	0.15
	FAR/(mm·h^-¹)	0.64	0.50	0.43	0.32	0.73
	CSI/(mm·h^-¹)	0.20	0.29	0.34	0.39	0.15

下载: 导出CSV

表 6 消融实验评价结果

评价指标	阈值	模型	3h	6h	9h	12h	15h	18h	21 h	24 h	平均
RMSE/(mm·h^-¹)	/	U-Net	2.500	2.645	2.801	2.953	3.102	3.251	3.398	3.549	2.977
		LU-Net	2.187	2.195	2.203	2.212	2.220	2.228	2.236	2.244	2.263
		GU-Net	0.902	1.501	1.799	2.050	2.301	2.450	2.600	2.751	1.914
		LGU-Net	0.903	0.911	0.904	0.908	0.912	0.905	0.910	0.906	0.912
MAE/(mm·h^-¹)	/	U-Net	1.452	1.583	1.754	1.923	2.134	2.345	2.568	2.788	2.067
		LU-Net	1.309	1.321	1.333	1.345	1.357	1.369	1.381	1.393	1.362
		GU-Net	0.612	0.724	0.856	1.012	1.165	1.321	1.489	1.654	1.147
		LGU-Net	0.638	0.645	0.634	0.631	0.637	0.642	0.639	0.635	0.633
POD/(mm·h^-¹)	0.1	U-Net	0.76	0.75	0.74	0.70	0.70	0.66	0.65	0.64	0.70
		LU-Net	0.76	0.77	0.79	0.76	0.77	0.78	0.76	0.76	0.77
		GU-Net	0.90	0.88	0.85	0.82	0.79	0.76	0.74	0.72	0.82
		LGU-Net	0.92	0.91	0.90	0.91	0.92	0.90	0.91	0.92	0.91
FAR/(mm·h^-¹)	0.1	U-Net	0.61	0.61	0.64	0.65	0.69	0.69	0.73	0.74	0.66
		LU-Net	0.60	0.61	0.60	0.62	0.58	0.60	0.61	0.60	0.60
		GU-Net	0.48	0.51	0.51	0.53	0.57	0.58	0.59	0.63	0.57
		LGU-Net	0.48	0.47	0.46	0.48	0.47	0.48	0.46	0.47	0.48
CSI/(mm·h^-¹)	0.1	U-Net	0.37	0.35	0.34	0.33	0.32	0.31	0.30	0.28	0.33
		LU-Net	0.36	0.37	0.38	0.37	0.36	0.37	0.38	0.37	0.37
		GU-Net	0.48	0.46	0.45	0.43	0.42	0.41	0.41	0.39	0.43
		LGU-Net	0.48	0.51	0.50	0.49	0.50	0.51	0.50	0.49	0.50

下载: 导出CSV

参考文献(23)

[1]	BAUER P, THORPE A, BRUNET G. The quiet revolution of numerical weather prediction[J]. Nature, 2015, 525(7 567): 47-55.
[2]	KLEIN W H, LEWIS B M, ENGER I. Objective prediction of five-day mean temperatures during winter[J]. J Atmos Sci, 2010, 16(6): 672- 682.
[3]	GLAHN H R, LOWRY D A. The use of Model Output Statistics (MOS) in objective weather forecasting[J]. J Appl Meteorol, 1972, 11(8): 1 203-1 211.
[4]	HOMLEID M. Diurnal corrections of short-term surface temperature forecasts using the Kalman filter[J]. Wea Forecasing, 1995, 10(4): 689- 707.
[5]	钱维宏. 如何提高天气预报和气候预测的技巧?[J]. 地球物理学报, 2012, 55(5): 1 532-1 540.
[6]	PENG X, CHE Y, CHANG J. A novel approach to improve numerical weather prediction skills by using anomaly integration and historical data[J]. J Geophys Res Atmos, 2013, 118(16): 8 814-8 826.
[7]	MARZBAN C. Neural networks for postprocessing model output: ARPS[J]. Mon Wea Rev, 2003, 131(6): 1 103-1 111.
[8]	RODRIGUES E R, OLIVEIRA I, CUNHA R L F, et al. DeepDownscale: A deep learning strategy for high-resolution weather forecast[J]. IEEE Comput Soc, 2018, 415-422.
[9]	SHI X J, CHEN Z R, WANG H, et al. Convolutional LSTM network: A machine learning approach for precipitation nowcasting[C]. Proceedings of the 29th International Conference on Neural Information Processing Systems, 2015, 1: 802-810.
[10]	郭瀚阳, 陈明轩, 韩雷, 等. 基于深度学习的强对流高分辨率临近预报试验[J]. 气象学报, 2019, 77(4): 715-727.
[11]	LEINONEN J, GUILLAUME A, YUAN T. Reconstruction of cloud vertical structure with a generative adversarial network[J]. Geophys Res Lett, 2019, 46(12): 7 035-7 044.
[12]	HAN L, CHEN M, CHEN K, et al. A deep learning method for bias correction of ECMWF 24-240 h forecasts[J]. Adv Atmos Sci, 2021, 38 (9): 1 444-1 459.
[13]	李紫丁. 基于深度学习的气象要素格点预报模式订正[D]. 长沙: 湖南师范大学, 2019.
[14]	DURAI V R, BHOWMIK S K R. Prediction of Indian summer monsoon in short to medium range time scale with high resolution global forecast system (GFS) T574 and T382[J]. Climate Dyn, 2014, 42(5/6): 1 527-1 551.
[15]	HERSBACH H, BELL B, BERRISFORD P, et al. The ERA5 global reanalysis[J]. Quart J Roy Meteor Soc, 2020, 146(730): 1 999-2 049.
[16]	王瑞. 基于深度学习的葵花8卫星资料反演降水[J]. 运筹与模糊学, 2023, 13(3): 1 710-1 719.
[17]	李彤, 王春月, 刘思言, 等. 吉林一号卫星在河湖"清四乱"中的应用[J]. 卫星应用, 2024(2): 21-27.
[18]	曲春梅, 朱瑞飞, 李恒宇, 等. 吉林一号卫星在林业中的应用[J]. 卫星应用, 2024(2): 14-20.
[19]	SUN J Y, ZHU R F, GONG J L, et al. Cross-comparison between Jilin-1GF03B and Sentinel-2 multi-spectral measurements and phenological monitors[J]. IEEE Access, 2024, 12: 43 540-43 551.
[20]	SHI W Z, CABALLERO J, FERENC H, et al. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2016: 1 874-1 883.
[21]	RONNEBERGER O, FISCHER P, BROX T. U-Net: Convolutional networks for biomedical image segmentation[M]. Springer International Publishing, 2015, 9 351: 234-241.
[22]	HOCHREITER S, SCHMIDHUBER J. Long short-term memory[J]. Neural Computation, 1997, 9(8): 1 735-1 780.
[23]	GAN X L, WANG L D, CHEN Q, et al. GAU-Net: U-Net based on global attention mechanism for brain tumor segmentation[J]. J Phys Conf Ser, 2021, 1 861(1): 012 041