Research on multi-model integrated precipitation forecast based on feed forward neural network

doi:10.11755/j.issn.1006-7639(2024)-01-0117

Abstract

Abstract:

In order to improve the accuracy of quantitative precipitation forecasting in Shandong Province, the deep feedforward neural network (DFNN) and the optimal threat score (TS) weight ensemble method for precipitation grading were used to study the multi-model ensemble precipitation forecasting. Four groups of DFNN (ES, EM, SM, ESM) deep learning models were obtained by using the 24-hour cumulative precipitation forecast of Global Numerical Prediction System of the European Centre for Medium-Range Weather Forecasts, the Shanghai Numerical Prediction Model System of the China Meteorological Administration and the Mesoscale Numerical Weather Prediction System of the China Meteorological Administration from April to September 2019 for supervised training, and the Mul-OTS (Multi-mode Optimal Threat Score) integrated model was established by using the optimal TS weight integration method of multi-model precipitation classification. The down-scale grid prediction was made by using the accumulated precipitation of each model for 24 h from April to September 2020, and the comparison test and case analysis of five integrated schemes were carried out. The results show that the average relative error was reduced by the five integrated schemes with different starting time and lead times. The ESM scheme was the best, and the Mul-OTS scheme was the worst. All the four groups of DFNN schemes improved the accuracy of sunny and rainy prediction, the ESM scheme was the best, and the Mul-OTS scheme was lower than the model forecast. The four groups of DFNN schemes all improved the TS and ETS scores of each precipitation grade, and the improvement amplitude of weak precipitation was greater than that of strong precipitation. The Mul-OTS scheme was a negative technique for the correction of small precipitation levels, and the correction effect was better for the correction of large precipitation levels, but it was still inferior to the ESM scheme. A case study found that the ESM scheme for precipitation intensity and fall area forecast was superior to other integrated schemes. Therefore, the optimal ESM scheme was adopted to establish a quantitative precipitation grid forecasting system, which provides important support for intelligent grid forecasting.

Key words: feedforward neural network, optimal threat score weight, multi-mode integration, grid precipitation forecast

摘要：

为提高山东定量降水预报准确率，采用深度前馈神经网络（Deep Forword Neural Networks， DFNN）和降水分级最优TS（Threat Score）权重集成方法对多模式集成降水预报进行研究。对2019年4—9月欧洲中期天气预报中心（European Centre for Medium-Range Weather Forecasting， ECMWF）全球数值预报系统、中国气象局上海数值预报模式系统（China Meteorological Administration Shanghai 9 km， CMA-SH9）和中国气象局中尺度天气数值预报系统（China Meteorological Administration Mesoscale， CMA-MESO）逐24 h累积降水量预报进行有监督训练，得到4组DFNN（ES、EM、SM、ESM）深度学习模型，并利用多模式降水分级最优TS权重集成方法建立Mul-OTS（Multi-mode Optimal Threat Score）集成模型。用2020年4—9月各模式逐24 h累积降水量进行降尺度格点预报，对5种集成方案对比检验、个例分析应用。结果表明：不同起报时间、不同预报时效，5组集成方案均降低了平均相对误差，ESM方案最好，Mul-OTS方案最差；4组DFNN方案均提高了晴雨准确率，ESM方案最好，Mul-OTS方案低于模式预报；4组DFNN方案均提高了各降水等级TS、ETS评分，对弱降水的提高幅度大于强降水，Mul-OTS方案对小量级降水等级订正是负技巧，对大量级降水等级的订正效果较好，但仍不如ESM方案；个例分析发现降水强度和落区预报ESM方案均优于其他集成方案。因此业务上采用最优的ESM方案建立了定量降水格点预报系统，为智能网格预报提供重要支撑。

关键词: 前馈神经网络, 最优TS权重, 多模式集成, 格点降水预报

CLC Number:

P457.6

ZHU Wengang, SHENG Chunyan, FAN Sudan, RONG Yanmin, QU Meihui. Research on multi-model integrated precipitation forecast based on feed forward neural network[J]. Journal of Arid Meteorology, 2024, 42(1): 117-128.

朱文刚, 盛春岩, 范苏丹, 荣艳敏, 曲美慧. 基于前馈神经网络的多模式集成降水预报研究[J]. 干旱气象, 2024, 42(1): 117-128.

Figures/Tables 12

Fig.1 The number of precipitation samples of different grades at 123 national stations in Shandong Province from April to September in 2019 and 2020

Fig.2 Threat score (a) and weight coefficients (b) of 24 h accumulated precipitation of different grades predicted by different models from April to September 2019

Fig.3 Structure diagram of deep feedforward neural network

Tab.1 Design scheme

方案	模式	简称
1	ECMWF、CMA-SH9	ES
2	ECMWF、CMA-MESO	EM
3	CMA-SH9、CMA-MESO	SM
4	ECMWF、CMA-SH9、CMA-MESO	ESM
5	ECMWF、CMA-SH9、CMA-MESO	Mul-OTS

Fig.4 The accuracy of sunny and rainy of 10 times cross validation using DFNN method and the model prediction with different lead times in Shandong Province from April to September 2019 (a) starting at 08:00, (b) starting at 20:00

Tab.2 The mean relative error of 10 times cross validation using DFNN method and model prediction with different start times and different lead times in Shandong Province from April to September 2019

预报时效	ECMWF		CMA-SH9		CMA-MESO		ES		EM		SM		ESM
预报时效	08：00 起报	20：00起报	08：00起报	20：00起报	08：00起报	20：00起报	08：00起报	20：00起报	08：00起报	20：00起报	08：00起报	20：00起报	08：00起报	20：00起报
0~24 h	0.51	0.52	0.59	0.59	0.57	0.57	0.47	0.47	0.47	0.46	0.48	0.49	0.46	0.45
24~48 h	0.58	0.57	0.60	0.60	0.65	0.66	0.49	0.48	0.50	0.52	0.52	0.51	0.49	0.48
48~72 h	0.63	0.62	0.65	0.65	0.69	0.70	0.52	0.52	0.53	0.54	0.55	0.54	0.51	0.52

Tab.3 The mean relative error of 24-hour cumulative precipitation predicted with different starting time and different lead times by different models and different schemes in Shandong Province from April to September 2020

预报时效	ECMWF		CMA-SH9		CMA-MESO		ES		EM		SM		ESM		Mul-OTS
预报时效	08：00 起报	20：00 起报	08：00起报	20：00起报	08：00起报	20：00起报	08：00起报	20：00起报	08：00起报	20：00起报	08：00起报	20：00起报	08：00起报	20：00起报	08：00起报	20：00起报
0~24 h	0.46	0.43	0.36	0.43	0.34	0.37	0.27	0.26	0.26	0.25	0.26	0.26	0.25	0.24	0.32	0.34
24~48 h	0.48	0.45	0.46	0.45	0.41	0.40	0.32	0.30	0.31	0.28	0.34	0.30	0.29	0.30	0.40	0.37
48~72 h	0.50	0.48	0.50	0.50	0.44	0.42	0.36	0.34	0.33	0.32	0.38	0.34	0.35	0.32	0.42	0.40

Fig.5 The accuracy of sunny and rainy of 24-hour cumulative precipitation predicted by different models and different schemes with different lead times in Shandong Province from April to September 2020 (a) starting at 08:00, (b) starting at 20:00

Fig.6 Threat score (a), miss rate (b), false alarm rate (c), bias score (d) and equitable threat score (e) of 0-24 h cumulative precipitation predicted by different models and different schemes in Shandong Province from April to September 2020

Fig.7 Threat score (a), miss rate (b), false alarm rate (c), bias score (d) and equitable threat score (e) of 24-48 h cumulative precipitation predicted by different models and different schemes in Shandong Province from April to September 2020

Fig.8 The observed values and the predicted values of different models and different schemes of 24 h cumulative precipitation from 08:00 on 23 to 08:00 on 24 August 2021 (a) observation, (b) ECMWF, (c) CMA-SH9, (d) CMA-MESO, (e) ES, (f) EM, (g) SM, (h) ESM, (i) Mul-OTS

Fig.9 The accuracy of sunny and rainy (a, b) and equitable threat score (c, d) of cumulative precipitation in 12~36 h (a, c) and 36~60 h (b, d) from April to September 2020 in Shandong Province were predicted by different models and different schemes

References 31

[1]	阿斯顿·张, 李沐, 扎卡里C.立顿, 等, 2019. 动手学深度学习[M]. 北京: 人民邮电出版社.
[2]	毕宝贵, 代刊, 王毅, 等, 2016. 定量降水预报技术进展[J]. 应用气象学报, 27(5): 534-549.
[3]	曹萍萍, 肖递祥, 龙柯吉, 等, 2023. 基于分位数映射法的四川省ECMWF模式降水预报误差订正分析[J]. 干旱气象, 41(4): 666-675. DOI
[4]	陈锦鹏, 冯业荣, 蒙伟光, 等, 2021. 基于卷积神经网络的逐时降水预报订正方法研究[J]. 气象, 47(1): 60-70.
[5]	杜钧, 康志明, 2014. 天气预报中不确定性问题的调查分析[J]. 气象科技进展, 4(1): 58-67.
[6]	顾建峰, 周国兵, 刘伯骏, 等, 2020. 人工智能技术在重庆临近预报业务中的初步研究与应用[J]. 气象, 46(10): 1 286-1 296.
[7]	李金洁, 王爱慧, 2019. 基于西南地区台站降雨资料空间插值方法的比较[J]. 气候与环境研究, 24(1): 50-60.
[8]	李莉, 李应林, 田华, 等, 2011. T213全球集合预报系统性误差订正研究[J]. 气象, 37(1): 31-38.
[9]	李俊, 杜钧, 陈超君, 2014. 降水偏差订正的频率(或面积)匹配方法介绍和分析[J]. 气象, 40(5): 580-588.
[10]	李俊, 杜钧, 陈超君, 2015. “频率匹配法”在集合降水预报中的应用研究[J]. 气象, 41(6): 674-684.
[11]	李涛, 陈杰, 汪方, 等, 2022. 一种基于神经网络的中国区域夏季降水预测订正算法[J]. 干旱气象, 40(2): 308-316. DOI
[12]	穆穆, 陈博宇, 周菲凡, 等, 2011. 气象预报的方法和不确定性[J]. 气象, 37(1): 1-13.
[13]	任宏利, 丑纪范, 2007. 数值模式的预报策略和方法研究进展[J]. 地球科学进展, 22(4): 376-385. DOI
[14]	孙健, 曹卓, 李恒, 等, 2021. 人工智能技术在数值天气预报中的应用[J]. 应用气象学报, 32(1): 1-11.
[15]	童尧, 高学杰, 韩振宇, 等, 2017. 基于RegCM4模式的中国区域日尺度降水模拟误差订正[J]. 大气科学, 41(6): 1 156-1 166.
[16]	吴启树, 韩美, 刘铭, 等, 2017. 基于评分最优化的模式降水预报订正算法对比[J]. 应用气象学报, 28(3): 306-317.
[17]	危国飞, 刘会军, 吴启树, 等, 2020. 多模式降水分级最优化权重集成预报技术[J]. 应用气象学报, 31(6): 668-680.
[18]	王海霞, 智协飞, 2015. 基于TIGGE多模式降水量预报的统计降尺度研究[J]. 气象科学, 35(4): 430-437.
[19]	王在文, 郑祚芳, 陈敏, 等, 2012. 支持向量机非线性回归方法的气象要素预报[J]. 应用气象学报, 23(5): 562-570.
[20]	王雨, 闫之辉, 2007. 降水检验方案变化对降水检验评估效果的影响分析[J]. 气象, 33(12): 53-61.
[21]	叶笃正, 严中伟, 戴新刚, 等, 2006. 未来的天气气候预测体系[J]. 气象, 32(4): 3-8.
[22]	袁凯, 庞晶, 李武阶, 等, 2023. 深度学习模型在2021年汛期武汉市雷达回波临近预报中的应用评估[J]. 干旱气象, 41(1): 173-185. DOI
[23]	袁雅涵, 王烁, 王文青, 等, 2023. 基于深度学习的积层混合云对流泡降水粒子特征研究[J]. 干旱气象, 41(6): 933-943.
[24]	周迪, 陈静, 陈朝平, 等, 2015. 暴雨集合预报-观测概率匹配订正法在四川盆地的应用研究[J]. 暴雨灾害, 34(2): 97-104.
[25]	智协飞, 吕游, 2019. 基于频率匹配法的中国降水多模式预报订正研究[J]. 大气科学学报, 42(6): 814-823.
[26]	智协飞, 赵忱, 2020. 基于集合成员订正的强降水多模式集成预报[J]. 应用气象学报, 31(3): 303-314.
[27]	智协飞, 张珂珺, 田烨, 等, 2021. 基于神经网络和地理信息的华东及华南地区降水概率预报[J]. 大气科学学报, 44(3): 381-393.
[28]	EBERT E E, 2001. Ability of a poor man’s ensemble to predict the probability and distribution of precipitation[J]. Monthly Weather Review, 129(10): 2 461-2 480.
[29]	HINTON G E, OSINDERO S, TEH Y W, 2006. A fast learning algorithm for deep belief nets[J]. Neural Computation, 18(7): 1 527-1 554.
[30]	JOLLIFFE I T, STEPHENSON D B, 2003. Forecast verification: A practitioner's guide in atmospheric science[M]. Chichester: John Wiley&Sons.
[31]	PANOFSKY H A, BRIER G H, 1968. Some applications of statistics to meteorology[M]. Philadelphia: The Pennsylvania State University Press:224-225.