﻿ 基于集合预报系统的日最高和最低气温预报
 气象学报  2017, Vol. 75 Issue (2): 211-222 PDF
http://dx.doi.org/10.11676/qxxb2017.023

0

文章信息

XIONG Minquan. 2017.

Calibrating daily 2 m maximum and minimum air temperature forecasts in the ensemble prediction system

Acta Meteorologica Sinica, 75(2): 211-222.
http://dx.doi.org/10.11676/qxxb2017.023

文章历史

2016-08-25 收稿
2016-12-26 改回

Calibrating daily 2 m maximum and minimum air temperature forecasts in the ensemble prediction system
XIONG Minquan
National Meteorological Center, Beijing 100081, China
Abstract: BP neural network-Self Memory method (BP-SM) is used to calibrate daily 2 m maximum (minimum) air temperature forecasts at 512 stations in China with the European Centre for Medium-Range Weather Forecasts (ECMWF) Ensemble Prediction System (EPS) in March 2016. Seven statistical characteristics used as predictors are calculated based on 2 m air temperature model output of EPS. Daily maximum (minimum) air temperature forecasts by BP-SM, BP and direct model output (DMO) are compared. The post-processing with BP-SM is shown to improve the forecast accuracy. Compared with BP method, more advantages of BP-SM method are attained in longer predictable time. The accurate rate of daily maximum (minimum) air temperature forecasts with absolute errors less than 2℃ reaches above 60% and even over 90% at some stations. Compared with DMO, the forecasting skill score of BP-SM is 30% on average, and above 60% over the eastern Tibetan Plateau. This program is obviously superior with forecast errors within 2℃(1℃). The calibrated daily 2 m maximum air temperature is slightly better than the daily 2 m minimum air temperature. By BP-SM method, the systematic deficiencies of daily 2 m maximum (minimum) air temperature forecasts are significantly reduced.
Key words: Ensemble prediction system     Daily maximum air temperature     Daily minimum air temperature     BP neural network-Self Memory method
1 引言

2 资料和方法 2.1 资料

 图 1 512个测站的分布 Figure 1 Geographic distribution of 512 stations in China
2.2 方法

BP (Back-Propagaion) 网络是人工神经网络的重要组成部分，广泛应用于诸多领域，其由3个环节构成，一是正向传播：信号从输入层经隐层到输出层；二为反向传播：预测误差从输出层，经原来的通道到达输入层；三是迭代过程：设立误差目标，当正向传播未达到目标值，通过反向传播，调整网络权值和阈值，反复迭代直至达到目标值。本研究使用集合预报的7个统计量作为输入层的7个节点，由Kolmogorov定理，隐层节点数设为15，输出层是1个节点，即预测和实际观测之差。构成3层网络进行日最高 (低) 气温预报，以预测误差 (观测值减DMO) 为目标值，出于实用考虑，迭代过程均为1000步，同时，网络学习过程采用经典的动量梯度法，步长为0.05。

 图 2 2015年12月21日至2016年2月29日08时上海南汇 (58369) DMO日最高气温24 h预报偏低的误差序列自相关函数 Figure 2 Autocorrelation function of error series for DMO underestimated daily 2 m maximum air temperature forecasts over Nanhui of Shanghai during the period of 21 Dec 2015-29 Feb 2016 at 08:00 BT
 (1)

 (2)

yt的估计值，

 (3)
 (4)
 (5)
 (6)

3 结果分析 3.1 BP-SM和BP方法的比较

 图 3 2016年3月DMO、BP和BP-SM方法的日最高气温预报比较 (a.1至10 d的BP-SM方法“正 (负) 效应”预报次数和误差增减值; b.上海南汇24 h预报误差的逐日比较 (主图)，1—5 d的预报误差和绝对误差月均值比较 (子图); c.拉萨市24 h预报误差的逐日比较 (主图)，1—5 d的预报误差和绝对误差月均值比较 (子图)) Figure 3 Comparisons of DMO, BP and BP-SM for monthly mean daily 2 m maximum air temperature forecasts during the period of 1-31 Mar 2016, (a) frequency, increment and decrement to forecast errors of positive and negative effects for BP-SM method over 10 d lead time; (b) comparisons (main diagram) of daily 24 h forecast errors over Nanhui of Shanghai, and monthly mean of daily 2 m minimum air temperature forecast errors and absolute errors (subgraph) over 5 d lead time; (c) comparisons (main diagram) of daily 24 h forecast errors over Lhasa and monthly mean of daily 2 m minimum air temperature forecast errors and absolute errors (subgraph) over 5 d lead time)

BP网络预测精度和数值模式预报能力密切相关。有较大误差的训练样本反馈在网络权值上，网络不稳定性增大，导致较大的输出偏差，表现为网络学习和记忆是不稳定的，有别于人类大脑记忆的稳定性，也是泛化性问题的反映；中长期 (2 d以上) 预报中，数值模式预报误差逐渐增大，BP-SM方法有较好的效果；对于短时 (1 d) 预报，BP方法预报有较高精度，需要更细致的分析，BP-SM方法才能达到预期。

3.2 绝对误差、准确率和技巧评分空间分布

 图 4 2016年3月BP-SM方法在1至5 d气温预报的月平均技巧评分 (a—e.最高气温, f—j.最低气温) Figure 4 Monthly mean skill-scores of temperature forecasts over China for the period of 1 Mar 2016-31 Mar 2016(a-e.daily 2 m maximum air temperature, f-j.daily 2 m minimum air temperature)

3.3 不同误差范围准确率的变化特点

 (a.日最高气温的正技巧比较, b.日最高气温的负技巧比较, c.日最低气温的正技巧比较, d.日最低气温的负技巧比较) 图 5 2016年3月DMO和BP-SM方法在不同误差范围的技巧评分比较 (a. positive skill-score of daily 2 m maximum air temperature forecasts, b. negative skill-score of daily 2 m maximum air temperature forecasts, c. positive skill-score of daily 2 m minimum air temperature forecasts, d. negative skill-score of daily 2 m minimum air temperature forecasts) Figure 5 The comparisons of skill-scores in different error ranges for DMO and BP-SM method during 1-31 Mar 2016

3.4 系统偏差比较

 (a.DMO和实况, b.BP-SM和实况, c.DMO和BP-SM比较) 图 6 2016年3月日最高气温24 h预报-实况散点图和不同预报误差范围站次数对比 (a. DMO and Obs, b. BP-SM and OBs, c. DMO and BP-SM) Figure 6 Forecast-observation scatterplots and frequencies of different error ranges for daily 2 m maximum air temperature forecasts at 24 h lead time during 1-31 Mar 2016

 (a.DMO和实况, b.BP-SM和实况) DMO和BP-SM方法的日最高气温24 h预报在不同误差范围站次数对比 (c.100°E以西地区，d.100°E以东地区) 图 7 2016年3月100°E以西和以东地区日最高气温24 h预报-实况散点图 (a) DMO, (b) BP-SM. Frequency comparisons between DMO and BP-SM methods in various error ranges over different areas:(c) to the west of 100°E, (d) to the east of 100°E Figure 7 Forecast-observation scatterplots for daily 2 m maximum temperature forecasts at 24 h lead time over the west and east of 100°E in the during 1-31 Mar 2016:

4 结论和讨论

(1) BP-SM和BP方法预报精度对比分析中：随着预报时效的延长，BP-SM方法优势也逐渐增大，“负效应”预报次数明显下降，表明BP-SM法能有效地减少BP网络预报异常值。上海南汇多个预报时效比较中：BP网络的不稳定性导致较大的预报误差；而BP-SM法避免了上述现象，在较长时效预报中，BP-SM法有更大的正向作用。通过上海南汇和拉萨两单站的DMO预报误差变化特点，可初步获悉EPS误差时空特点和BP网络预报准确度有密切关系。以上海南汇为例，较细致地介绍了BP-SM方法日常预报过程。

(2) BP-SM方法十分显著地提高了日最高 (低) 气温预报准确率，尤其是在青藏高原东部和南部地区。和DMO相比，BP-SM方法在大部分地区都有较高的正技巧，部分地区超过了60%，负技巧站点通常出现在DMO预报准确率较高的部分地区，技巧分值较小 (0—-20%)。

(3) 具体比较了绝对误差≤2℃(1℃) 准确率变化特点，在未来5 d预报中，BP-SM方法在此误差范围内的正技巧站点比例表现出更大的优势，同时，对日最高气温预报准确率的提高能力要略好于日最低气温。探讨了BP-SM方法出现负技巧预报的原因，说明BP-SM方法有较强的非线性映射能力，当面对函数逼近精度和模型容错性等问题时，还要与天气变化特点相联系，可能会达到更好的预报效果。

(4) 实况-预报散点图及在东、西部站点比较中，DMO呈现发散、非对称性分布特点，而BP-SM方法减小了系统性偏差，数值点也收敛、对称地分布。BP-SM方法也改变了DMO在6个误差区段分布形态的不一致性，使得误差越大区段站点数越少，分布结构呈快速衰减趋势。

 陈超辉, 李崇银, 谭言科, 等. 2010. 基于交叉验证的多模式超级集合预报方法研究. 气象学报, 68(4): 464–476. Chen C H, Li C Y, Tan Y K, et al. 2010. Research of the multi-model super-ensemble prediction based on cross-validation. Acta Meteor Sinica, 68(4): 464–476. DOI:10.11898/1001-7313.20100410 (in Chinese) 杜钧, GrummR H, 邓国. 2014. 预报异常极端高影响天气的"集合异常预报法":以北京2012年7月21日特大暴雨为例. 大气科学, 38(4): 685–699. Du J, Grumm R H, Deng G. 2014. Ensemble anomaly forecasting approach to predicting extreme weather demonstrated by extremely heavy rain event in Beijing. Chinese J Atmos Sci, 38(4): 685–699. DOI:10.3878/j.issn.1006-9895.2013.13218 (in Chinese) 刘建国, 谢正辉, 赵琳娜, 等. 2013. 基于TIGGE多模式集合的24小时气温BMA概率预报. 大气科学, 37(1): 43–53. Liu J G, Xie Z H, Zhao L N, et al. 2013. BMA probabilistic forecasting for the 24-h TIGGE multi-model ensemble forecasts of surface air temperature. Chinese J Atmos Sci, 37(1): 43–53. DOI:10.3878/j.issn.1006-9895.2012.11232 (in Chinese) 刘琳, 陈静, 程龙, 等. 2011. 基于集合预报的中国极端强降水预报方法研究. 气象学报, 71(5): 853–866. Liu L, Chen J, Cheng L. 2011. Study of the ensemble-based forecast of extremely heavy rainfalls in China: Experiments for July 2011 cases. Acta Meteor Sinica, 71(5): 853–866. (in Chinese) 刘永和, 严中伟, 冯锦明, 等. 2013. 基于TIGGE资料的沂沭河流域6小时降水集合预报能力分析. 大气科学, 37(3): 539–551. Liu Y H, Yan Z W, Feng J M, et al. 2013. Predictability of 6-hour precipitation in the Yishu River basin based on TIGGE data. Chinese J Atmos Sci, 37(3): 539–551. DOI:10.3878/j.issn.1006-9895.2012.11078 (in Chinese) 马旭林, 时洋, 和杰, 等. 2015. 基于卡尔曼滤波递减平均算法的集合预报综合偏差订正. 气象学报, 73(5): 952–964. Ma X L, Shi Y, He J, et al. 2015. The combined descending averaging bias correction based on the kalman filter for ensemble forecast. Acta Meteor Sinica, 73(5): 952–964. (in Chinese) 王敏, 李晓莉, 范广洲, 等. 2012. 区域集合预报系统2 m温度预报的校准技术. 应用气象学报, 23(4): 395–401. Wang M, Li X L, Fan G Z, et al. 2012. Calibrating 2 m temperature forecast for the regional ensemble prediction system at NMC. J Appl Meteor Sci, 23(4): 395–401. DOI:10.11898/1001-7313.20120402 (in Chinese) Gneiting T, Raftery A E, Westveld A H, et al. 2005. Calibrated probabilistic forecasting using ensemble model output statistics and minimum CRPS estimation. Mon Wea Rev, 133(5): 1098–1118. DOI:10.1175/MWR2904.1 Hamill T M, Colucci S J. 1997. Verification of Eta-RSM short-range ensemble forecasts. Mon Wea Rev, 125(6): 1312–1327. DOI:10.1175/1520-0493(1997)125<1312:VOERSR>2.0.CO;2 Hamill T M, Colucci S J. 1998. Evaluation of Eta-RSM ensemble probabilistic precipitation forecasts. Mon Wea Rev, 126(3): 711–724. DOI:10.1175/1520-0493(1998)126<0711:EOEREP>2.0.CO;2 Hamill T M, Whitaker J S, Wei X. 2004. Ensemble reforecasting: Improving medium-range forecast skill using retrospective forecasts. Mon Wea Rev, 132(6): 1434–1447. DOI:10.1175/1520-0493(2004)132<1434:ERIMFS>2.0.CO;2 Hamill T M, Whitaker J S, Mullen S L. 2006. Reforecasts: An important dataset for improving weather predictions. Bull Amer Meteor Soc, 87(1): 33–46. DOI:10.1175/BAMS-87-1-33 Mendoza P A, Rajagopalan B, Clark M P, et al. 2015. Statistical postprocessing of high-resolution regional climate model output. Mon Wea Rev, 143(5): 1533–1553. DOI:10.1175/MWR-D-14-00159.1 Raftery A E, Gneiting T, Balabdaoui F, et al. 2005. Using Bayesian model averaging to calibrate forecast ensembles. Mon Wea Rev, 133(5): 1155–1174. DOI:10.1175/MWR2906.1 Roulin E, Vannitsem S. 2012. Postprocessing of ensemble precipitation predictions with extended logistic regression based on hindcasts. Mon Wea Rev, 140(3): 874–888. DOI:10.1175/MWR-D-11-00062.1 Roulston M S, Smith L A. 2003. Combining dynamical and statistical ensembles. Tellus A, 55(1): 16–30. DOI:10.3402/tellusa.v55i1.12082 Schmeits M J, Kok K J. 2010. A comparison between raw ensemble output, (modified) Bayesian model averaging, and extended logistic regression using ECMWF ensemble precipitation reforecasts. Mon Wea Rev, 138(11): 4199–4211. DOI:10.1175/2010MWR3285.1 Stephenson D B, Coelho C A S, Doblas-Reyes F J, et al. 2005. Forecast assimilation: A unified framework for the combination of multi-model weather and climate predictions. Tellus A, 57(3): 253–264. DOI:10.3402/tellusa.v57i3.14664 Wilks D S, Hamill T M. 2007. Comparison of ensemble-MOS methods using GFS reforecasts. Mon Wea Rev, 135(6): 2379–2390. DOI:10.1175/MWR3402.1