周康辉，郑永光，王婷波. 2021. 利用深度学习融合NWP和多源观测数据的闪电落区短时预报方法. 气象学报，79（1）：1-14 doi: 10.11676/qxxb2021.002
引用本文:
周康辉，郑永光，王婷波. 2021. 利用深度学习融合NWP和多源观测数据的闪电落区短时预报方法. 气象学报，79（1）：1-14
doi:
10.11676/qxxb2021.002
Zhou Kanghui， Zheng Yongguang， Wang Tingbo. 2021. Very short-range lightning forecasting with NWP and observation data: A deep learning approach. Acta Meteorologica Sinica， 79（1）:1-14 doi: 10.11676/qxxb2021.002
Citation:
Zhou Kanghui， Zheng Yongguang， Wang Tingbo. 2021. Very short-range lightning forecasting with NWP and observation data: A deep learning approach. Acta Meteorologica Sinica， 79（1）:1-14
doi:
10.11676/qxxb2021.002
周康辉，郑永光，王婷波. 2021. 利用深度学习融合NWP和多源观测数据的闪电落区短时预报方法. 气象学报，79（1）：1-14 doi: 10.11676/qxxb2021.002
引用本文:
周康辉，郑永光，王婷波. 2021. 利用深度学习融合NWP和多源观测数据的闪电落区短时预报方法. 气象学报，79（1）：1-14
doi:
10.11676/qxxb2021.002
Zhou Kanghui， Zheng Yongguang， Wang Tingbo. 2021. Very short-range lightning forecasting with NWP and observation data: A deep learning approach. Acta Meteorologica Sinica， 79（1）:1-14 doi: 10.11676/qxxb2021.002
Citation:
Zhou Kanghui， Zheng Yongguang， Wang Tingbo. 2021. Very short-range lightning forecasting with NWP and observation data: A deep learning approach. Acta Meteorologica Sinica， 79（1）:1-14
doi:
10.11676/qxxb2021.002
强对流短时预报（2—6 h）具有较大难度。一方面，基于观测数据的外推已基本不可用；另一方面，高分辨率数值模式（High-resolution Numerical Weather Prediction，HNWP）的预报性能有待提升。利用深度学习方法，将卫星、雷达、云-地闪电（简称闪电）等观测数据和高分辨率数值模式预测数据进行融合，得到更有效的闪电落区短时预报结果。基于多源观测数据和高分辨率数值天气预报数据的特性，构建了一个双输入单输出的深度学习语义分割模型（LightningNet-NWP），使用了包括闪电密度、雷达组合反射率拼图、卫星成像仪6个红外通道，以及GRAPES_3km模式预报的雷达组合反射率等共9个预报因子。深度学习模型使用了编码-解码的经典全卷卷积结构，并使用池化索引共享的方式，尽可能保留不同尺度特征图上的细节特征信息；利用三维卷积层提取观测数据时间和空间上的变化特征。结果表明，LightningNet-NWP能够较好地实现0—6 h的闪电落区预报，具备比单纯使用多源观测数据、高分辨率数值模式预报数据更好的预报结果。深度学习能够有效实现多源观测数据和数值天气预报数据的融合，在2—6 h时效预报效果优于单独使用观测数据或数值天气预报数据；预报时效越长，融合的优势体现得越明显。
强对流 /
短时预报 /
深度学习 /
观测数据 /
数值模式预报
Abstract:
The very short-range (VSR, 2—6 h) convective weather forecasting is still a great challenge. On the one hand, the extrapolation of observation data is no longer available. On the other hand, the High-resolution Numerical Weather Prediction (HNWP) performance needs to be further improved. To address the above issues, a semantic segmentation deep learning network named LightningNet-NWP is implemented to merge the multi-source observation data with HNWP data to get better VSR lightning forecasts. The predictors of the LightningNet-NWP include lightning density, radar reflectivity, 6 infrared bands of Himawari-8 and the radar composite reflectivity from GRAPES_3km. Because the observations and HNWP data differ a lot, two encode-decode symmetry sub-networks were designed to extract future information from the above two data sources. The pooling index is shared in upsampling process, so that the details of shallow feature maps are transmitted and fully used. Three dimensional convolutional layers are utilized to extract spatial and temporal features. The experimental results show that the LightningNet-NWP can effectively combine observations and HNWP data and yield a good lightning prediction for the next 0—6 hours. The performance of the LightningNet-NWP combined with observations and HNWP data is much better than that solely using observations or HNWP data. The longer the prediction period, the more advantageous the combinational use of observations and HNWP data.
Key words:
Convective weather /
Short-term forecast /
Deep learning /
Observation data /
Numerical weather prediction
Figure 5.
Predictions of Lightning-NWP （a
1
—a
6
） and Obs_Network （b
1
—b
6
） of the next 0—6 h and lightning observations on 23 August 2018 （a
1
，b
1
. 12：00—13:00 BT；a
2
，b
2
. 13：00—14:00 BT；a
3
，b
3
. 14：00—15:00 BT；a
4
，b
4
. 15:00—16:00 BT；a
5
，b
5
. 16：00—17:00 BT；a
6
，b
6
. 17：00—18：00 BT；contours are predictions and the black points denote corresponding lightning observations）
The performance of DL with different data sources，optical flow extrapolation with lightning data，and HNWP's output for convective weather （The lightning observation data of the previous 2 hours are used by optical flow method to forecast lightning activities of the next 0—6 hours.
h
，
m
，
f
，and
c
indicate the hits，misses，false alarms，and correct negatives，respectively，and
$ {h_{{\rm{random}}}} = (h + f) \times (h + m)/(h + m + f + c)$
）
方法时效PODFARBiasETSTS$ \dfrac{h}{h+m} $$\dfrac{f}{h+f} $$ \dfrac{h+f}{h+m} $$\dfrac{h-{h}_{\rm{random}}}{h+m+f-{h}_{\rm{random}}} $$ \dfrac{h}{h+m+f} $
深
度
学
习
方
法观测数据
和高分辨率
数值模式1 h0.6180.3891.0120.4010.4442 h0.5500.5301.1690.2950.3403 h0.5240.6341.4320.2230.2754 h0.4840.6821.5230.1840.2385 h0.490.7381.8730.1500.2066 h0.4630.7461.8260.1450.196观测数据1 h0.6330.3861.0310.4160.4532 h0.5330.5581.2060.2770.3193 h0.4570.6521.3120.2000.2464 h0.4290.7331.6100.1450.1975 h0.4020.7831.8500.1110.1646 h0.3720.8242.1120.0830.136高分辨率
数值模式1 h0.5010.7822.3000.1170.1792 h0.5370.7952.6120.1080.1753 h0.4730.7922.2290.1070.1734 h0.4720.7912.2540.1060.1705 h0.4740.7902.2550.1100.1716 h0.4570.8032.3180.1030.160基于闪电数据的
光流法外推预报1 h0.4130.4750.7870.2690.3012 h0.2490.6450.7010.1380.1713 h0.1860.7360.5670.0880.1234 h0.1260.8100.6610.0470.0825 h0.0940.8600.6760.0250.0606 h0.0760.8940.7170.0130.046GRAPES_3km
（40 dBz超过0℃层）0—6 h0.2910.8622.3780.0980.123
地址：北京市海淀区中关村南大街46号
邮政编码：100081
Email：
qxxb@cms1924.org
;
cmsqxxb@263.net
电话：010-68406942; 010-68408571
