基于SCTP-RF算法的甘肃省短期定量降水客观预报方法研究

doi:10.11755/j.issn.1006-7639(2022)-01-0146

摘要/Abstract

摘要：

基于欧洲中期天气预报中心(ECMWF)的精细化数值预报产品、中国气象局下发的降水指导产品(TP_CMA)及甘肃省340个气象站点降水实况数据,利用泰森多边形与K-means空间聚类方法(spatial cluster and Tyson polygon,SCTP),对2017—2019年4—9月甘肃省340站降水资料进行客观分区。在此基础上,采用随机森林算法(random forest,RF),筛选出与降水相关的物理量因子构建模型,开展甘肃省短期定量降水客观预报订正试验,并进行预报效果检验。结果表明:(1)甘肃省4—9月降水客观分区依次为7、6、14、13、14和11个。(2)就晴雨预报而言,SCTP-RF订正产品对甘肃省汛期的晴雨预报能力较TP_CMA指导产品和ECMWF模式产品有一定提升,提升幅度分别为6.1%、4.2%;在空间上,SCTP-RF算法对甘肃省340站的晴雨预报均具有一定的订正能力,大部分站点晴雨预报准确率提升了5%,特别是河东地区。(3)在分级降水预报中,SCTP-RF订正产品对中雨和大雨的预报能力均优于TP_CMA指导产品和ECMWF模式产品,且全省大部的订正效果较好,特别是河东中部及陇东南地区,但在强降水过程中对小雨和暴雨的预报订正不稳定,尤其是陇东南地区的小雨。

关键词: 降水及分级, 短期预报, 空间聚类, 随机森林, 预报订正及检验

Abstract:

Based on the refined numerical prediction products of the European Center of Medium-range Weather Forecast (ECMWF), precipitation guidance products from China Meteorological Administration (TP_CMA) and precipitation observation data at 340 meteorological stations of Gansu Province, the objective divisions of precipitation at 340 meteorological stations of Gansu Province from April to September during 2017-2019 were done by using spatial cluster and Tyson polygon (SCTP) approach. On this basis, the physical quantity factors related to precipitation were selected and used to build prediction model by using random forest (RF) algorithm, and the correction experiment of short-term quantitative precipitation objective forecast in Gansu Province was carried out, the forecast effect was verified. The results are as follows: (1) There were 7, 6, 14, 13, 14 and 11 precipitation regions in sequence from April to September in Gansu Province. (2) In terms of rain probability forecast, the forecast ability of SCTP-RF products in flood season (from June to August) in Gansu Province was better than that of TP_CMA guidance products and ECMWF model products, and the prediction accuracy improved by 6.1% and 4.2%, respectively. In space, SCTP-RF products had a certain ability to correct rain probability forecast at all stations of Gansu Province, and the prediction accuracy at most stations improved by 5%, especially in the east of Yellow River in Gansu (Hedong area). (3) For graded precipitation forecast, the forecast ability of SCTP-RF products to moderate rain and heavy rain was superior to TP_CMA guidance products and ECMWF model products, and the correction effect at most stations was better, especially in the middle part of Hedong and southeastern Gansu. However, the correction ability to light rain and rainstorm forecast was unstable during the heavy rainfall processes, especially to light rain in southeastern Gansu.

Key words: precipitation and classification, short-term forecast, spatial clustering, random forest, forecast correction and test

中图分类号:

P456.7
P45

刘娜, 黄武斌, 杨建才, 王基鑫, 王一丞, 张君霞. 基于SCTP-RF算法的甘肃省短期定量降水客观预报方法研究[J]. 干旱气象, 2022, 40(1): 146-155.

LIU Na, HUANG Wubin, YANG Jiancai, WANG Jixin, WANG Yicheng, ZHANG Junxia. Objective forecast method of short-term quantitative precipitation in Gansu Province based on SCTP-RF algorithm[J]. Journal of Arid Meteorology, 2022, 40(1): 146-155.

图/表 11

图1 随机森林算法示意图

Fig.1 Schematic map of random forest algorithm

表1 降水分级的检验评定

Tab.1 The inspection and assessment of precipitation grades

实况	预报
实况	小雨	中雨	大雨	暴雨及以上
小雨	NA₁	NB₂	NB₃	NB₄
中雨	NC₂	NA₂	NB₃	NB₄
大雨	NC₃	NC₃	NA₃	NB₄
暴雨及以上	NC₄	NC₄	NC₄	NA₄

图2 甘肃省4—9月降水客观分区 (a)4月,(b)5月,(c)6月,(d)7月,(e)8月,(f)9月

Fig.2 The objective partitions of precipitation in Gansu Province from April to September (a) April, (b) May, (c) June, (d) July, (e) August, (f) September

图3 定量降水预报的客观订正方案流程

Fig.3 The objective correction scheme flow of quantitative precipitation forecast

表2 调整后对降水预报模型有正贡献的因子及其贡献率单位:%

Tab.2 The positive contribution factors to precipitation forecast model after the adjustment and their contribution rates

因子	贡献率	因子	贡献率	因子	贡献率	因子	贡献率	因子	贡献率
TP_CMA	10.40	RH₂₀₀	5.40	RH₁₀₀	1.92	GH₅₀₀	0.62	GH₇₀₀	0.24
R	10.21	Q₅₀₀	4.71	CAPE	1.75	UV₅₀₀	0.58	T₁₀₀	0.16
△H₂₀₀	10.09	T₇₀₀	4.30	△H₅₀₀	1.34	GH₂₀₀	0.53	T₅₀₀	0.10
△H₇₀₀	8.61	$T d 2$	4.12	Q₇₀₀	1.27	UV₁₀₀	0.52	w₁₀₀	0.06
R_{o-3 h}	7.61	w₅₀₀	3.66	△H₇₀₀	1.27	GH₁₀₀	0.50	T₂	0.05
RH₅₀₀	7.18	w₇₀₀	2.35	△T₂₀₀	0.83	UV₂₀₀	0.40
RH₇₀₀	5.93	w₂₀₀	2.28	△T₅₀₀	0.63	T₂₀₀	0.38

表2 调整后对降水预报模型有正贡献的因子及其贡献率单位:%

Tab.2 The positive contribution factors to precipitation forecast model after the adjustment and their contribution rates

因子	贡献率	因子	贡献率	因子	贡献率	因子	贡献率	因子	贡献率
TP_CMA	10.40	RH₂₀₀	5.40	RH₁₀₀	1.92	GH₅₀₀	0.62	GH₇₀₀	0.24
R	10.21	Q₅₀₀	4.71	CAPE	1.75	UV₅₀₀	0.58	T₁₀₀	0.16
△H₂₀₀	10.09	T₇₀₀	4.30	△H₅₀₀	1.34	GH₂₀₀	0.53	T₅₀₀	0.10
△H₇₀₀	8.61	$T d 2$	4.12	Q₇₀₀	1.27	UV₁₀₀	0.52	w₁₀₀	0.06
R_{o-3 h}	7.61	w₅₀₀	3.66	△H₇₀₀	1.27	GH₁₀₀	0.50	T₂	0.05
RH₅₀₀	7.18	w₇₀₀	2.35	△T₂₀₀	0.83	UV₂₀₀	0.40
RH₇₀₀	5.93	w₂₀₀	2.28	△T₅₀₀	0.63	T₂₀₀	0.38

表3 2020年6—8月TP_CMA、ECMWF及SCTP-RF产品不同起报时次下逐3 h降水的晴雨预报准确率单位:%

Tab.3 The accuracy of 3-hour rain probability forecast by TP_CMA, ECMWF and SCTP-RF products at different start predicting times from June to August 2020

降水预报产品	08:00起报			20:00起报
降水预报产品	6月	7月	8月	6月	7月	8月
TP_CMA	79.29	78.98	74.87	76.94	77.46	75.33
ECMWF	78.62	75.73	83.34	77.64	75.71	83.25
SCTP-RF	84.88	82.94	82.76	83.57	82.74	82.60

图4 2020年6—8月08:00(a)、20:00(b)起报的ECMWF和SCTP-RF产品逐3 h降水的晴雨预报技巧评分

Fig.4 The skill score of 3-hour rain probability forecast by ECMWF and SCTP-RF products at 08:00 BST (a) and 20:00 BST (b) start predicting from June to August 2020

图5 2020年6—8月08:00(a、c、e)和20:00(b、d、f)起报的TP_CMA(a、b)、ECMWF(c、d)和SCTP-RF(e、f)产品晴雨预报准确率空间分布(单位:%)

Fig.5 The accuracy spatial distribution of rain probability forecast by TP_CMA (a, b), ECMWF (c, d) and SCTP-RF (e, f) products at 08:00 BST (a, c, e) and 20:00 BST (b, d, f) start predicting from June to August 2020 (Unit: %)

表4 2020年6—8月TP_CMA、ECMWF和SCTP-RF产品不同起报时次下逐3 h降水分级预报TS评分单位:%

Tab.4 The threat scores of 3-hour precipitation with different grades of TP_CMA, ECMWF and SCTP-RF products at different start predicting times from June to August 2020

降水等级	降水预报产品	6月		7月		8月
降水等级	降水预报产品	08:00起报	20:00起报	08:00起报	20:00起报	08:00起报	20:00起报
小雨	TP_CMA	18.54	17.48	18.56	19.41	20.96	21.13
	ECMWF	19.47	18.82	20.37	20.63	9.17	8.49
	SCTP-RF	20.81	18.94	19.29	20.50	13.69	13.68
中雨	TP_CMA	12.09	10.96	14.72	13.22	18.47	17.79
	ECMWF	15.44	16.28	16.75	17.00	21.05	19.94
	SCTP-RF	15.25	15.74	17.76	17.24	21.59	20.29
大雨	TP_CMA	7.97	6.67	6.84	6.33	10.00	12.08
	ECMWF	7.59	10.87	5.73	7.09	12.57	11.81
	SCTP-RF	8.74	10.61	7.32	8.39	12.24	12.88
暴雨及以上	TP_CMA	3.84	2.74	1.55	2.10	5.71	8.86
	ECMWF	5.51	10.06	2.89	3.07	8.15	6.40
	SCTP-RF	4.67	4.84	2.07	1.54	8.21	8.50

图6 2020年6—8月08:00(a、c、e、g)和20:00(b、d、f、h)起报的SCTP-RF产品逐3 h降水分级技巧评分空间分布(单位:%) (a、b)小雨,(c、d)中雨,(e、f)大雨,(g、h)暴雨

Fig.6 The spatial distribution of skill score of 3-hour precipitation with different grades of SCTP-RF products at 08:00 BST (a, c, e, g) and 20:00 BST (b, d, f, h) start predicting from June to August 2020 (Unit: %) (a, b) light rain, (c, d) moderate rain, (e, f) heavy rain, (g, h) rainstorm

图7 2020年8月TP_CMA、ECMWF和SCTP-RF产品逐3 h降水的小雨漏报率(a)和空报率(b)

Fig.7 Missing ratio (a) and false alarm ratio (b) of light rain of TP_CMA, ECMWF and SCTP-RF 3-hour precipitation products in August 2020

参考文献 28

[1]	宗志平, 代刊, 蒋星. 定量降水预报技术研究进展[J]. 气象科技进展, 2012, 2(5):29-35.
[2]	马振锋, 谭友邦. 预测川渝地区汛期降水量的一种物理统计模型[J]. 大气科学, 2004, 28(1):138-145.
[3]	智协飞, 季晓东, 张璟, 等. 基于TIGGE资料的地面气温和降水的多模式集成预报[J]. 大气科学学报, 2013, 36(3):257-266.
[4]	李巧, 戚友存, 朱自伟, 等. 复杂地形下C波段雷达定量降水估计算法[J]. 气象学报, 2021, 79(4):689-702.
[5]	苏华英, 曾莉萍, 罗乃兴, 等. 基于雷达资料的定量降水预报外推算法试验[J]. 沙漠与绿洲气象, 2020, 14(5):27-35.
[6]	MARKUS R, CAMPS-VALLS G, STEVENS B, et al. Deep learning and process understanding for data-driven Earth system science[J]. Nature, 2019, 566(7743):195-204. DOI URL
[7]	VALVERDE RAMÍREZ M C, DE CAMPOS VELHO H F, FERREIRA N J. Artificial neural network technique for rainfall forecasting applied to the São Paulo region[J]. Journal of Hydrology, 2005, 301:146-162. DOI URL
[8]	ANDREW K, WEI X P, PRAKASH V A, et al. Modeling and prediction of rainfall using radar reflectivity data: a data-mining approach[J]. IEEE Transactions on Geoscience and Remote Sensing, 2013, 51(4):2337-2342. DOI URL
[9]	朱岩, 翟丹华, 吴志鹏, 等. 基于Xgboost算法的短时强降水预报方法[J]. 气象科技, 2021, 49(3):406-418.
[10]	赵华生, 金龙, 黄小燕, 等. 基于CNN和RF算法的ECMWF降水分级订正预报方法[J]. 气象科技, 2021, 49(3):419-426.
[11]	孙日丁, 周官辉, 杜滨鹤, 等. 人工神经网络方法在鹤壁汛期降水预报中的应用[J]. 河南气象, 2004(3):12-13.
[12]	王伟健, 姚展予, 贾烁, 等. 随机森林算法在人工增雨效果统计检验中的应用研究[J]. 气象与环境科学, 2018, 41(2):111-117.
[13]	黄玺玮, 王利, 杨双宁, 等. 基于随机森林的内江市降水量中长期预测[J]. 阴山学刊(自然科学版), 2017, 31(4):107-110.
[14]	史静, 朱虹. 基于随机森林的天气场景判别算法[J]. 微型机与应用, 2017, 36(24):51-53.
[15]	甄亿位, 郝敏, 陆宝宏, 等. 基于随机森林的中长期降水量预测模型研究[J]. 水电能源科学, 2015, 33(6):6-10.
[16]	余胜男, 陈元芳, 顾圣华, 等. 随机森林在降水量长期预报中的应用[J]. 南水北调与水利科技, 2016, 14(1):78-83.
[17]	张晨阳, 杨雪冰, 张文生. 气象大数据超短临精准降水机器学习与典型应用[J]. 农业大数据学报, 2019, 1(1):78-87.
[18]	赵琳娜, 董航宇, 吴亮, 等. 黄淮地区夏季日降水分区概率预报方法研究[J]. 气象, 2015, 41(12):1503-1513.
[19]	赵翠光, 赵声蓉. 华北及周边地区夏季分区客观降水预报[J]. 应用气象学报, 2011, 22(5):558-566.
[20]	马亚丽, 敖天其, 王汉涛. 泰森多边形距离反比加权法在BTOPMC模型中的应用分析[J]. 水利规划与设计, 2019(11):58-63.
[21]	刘新伟, 蒋盈沙, 黄武斌, 等. 基于雷达产品和随机森林算法的冰雹天气分类识别及预报[J]. 高原气象, 2021, 40(4):898-908. DOI
[22]	方匡南, 吴见彬, 朱建平, 等. 随机森林方法研究综述[J]. 统计与信息论坛, 2011, 26(3):32-38.
[23]	唐亘. 精通数据科学: 从线性回归到深度学习[M]. 北京: 人民邮电出版社, 2018:238-239.
[24]	董师师, 黄哲学. 随机森林理论浅析[J]. 集成技术, 2013, 2(1):1-7.
[25]	王红霞, 柳小妮, 李纯斌, 等. 甘肃省近42年降水量变化时空分布格局分析[J]. 中国农业气象, 2013, 34(4):384-389.
[26]	刘峰民, 白冰, 赵小娟. 基于DEM的甘肃降水空间分布[J]. 甘肃科技, 2020, 36(15):50-53.
[27]	朱新胜, 张耀存. 次网格地形坡度坡向参数化及其对区域气候模拟的影响[J]. 高原气象, 2005, 24(2):136-142.
[28]	刘达, 张玲. 2020年8月大气环流和天气分析[J]. 气象, 2020, 46(11):1520-1528.