Page 84 - 《高原气象》2025年第3期

高 原 气 象 44 卷
642
Yu J W， et al， 2019. Impact of topographic altitude bias of the of Atmospheric Sciences， 42（5）： 652-659. DOI： 10. 13878/j.
GFS model on the 2 m air temperature forecast［J］. Transactions cnki. dqkxxb. 20180323001.

Analysis on the Error Correction Method of 2m Temperature
Hourly Forecast Based on CMA-GD Model

LI Jian ， FAN Qi ， ZHANG Ying ， XU Xingsheng 3
1， 2
3
1
（1. Sun Yat-sen University School of Atmospheric Sciences， Zhuhai 519082， Guangdong， China；
2. Xinyu Meteorological Bureau of Jiangxi Province， Xinyu 338000， Jiangxi， China；
3. Jiangxi Meteorological Observatory， Nanchang 330096， Jiangxi， China）

Abstract： The most significant meteorological component is temperature， and weather forecasting relies heavily
on how accurately temperatures are predicted. This study uses a linear non-graded regression method to rectify
the inaccuracies in temperature forecasts induced by terrain variation in the 2 m temperature hourly forecast prod‐
uct of the mesoscale numerical model （China Meteorological Administration Guangdong， CMA-GD）， and use
the one-dimensional Kalman filtering method and the double-weighted moving average method to correct the re‐
sults. The accuracy of the hourly distribution exhibits a diurnal variation feature， and the model terrain height de‐
viation is linearly negatively connected with the temperature error mean value， according to the results. The day‐
time correction impact is superior than the nighttime correction effect following the ungraded regression method.
recorrecting using the best time frame for mathematical correction techniques （15 days for the Kalman method
and 20 days for the mean method）. It is discovered that the mean method's re-correction effect outperforms the
Kalman methods， and that the correction effect is more pronounced during the day than at night. Summer and au‐
tumn have a better re-correction impact than winter and spring， with some negative correction effects in spring
and little difference between the two techniques in the latter. In the former， the mean value method outperforms
the Kalman method. There are eight stations with negative correction following the ungraded regression method，
but no negative correction stations follow the mathematical correction methods. Therefore the northern region
typically experiences a better corrective impact than the southern region. The fraction of correction magnitude for
both MAE and ACC is positively correlated with a binomial connection. The terrain deviation correction method
has the least slope and restricted correction effect， while the mean value approach has the best correlation and
largest slope. An error assessment was conducted in the middle part of Poyang Lake Plain and the south Zhejiang-
Fujian hilly region. The peak error value in the former was lower than that in the latter， and the correction ampli‐
tude at the peak was smaller. After correction， the MAE decreased by 25. 1% and 19. 8%， respectively. From No‐
vember 2022 to January 2023， during frequent cold air intrusions， the MAE in the middle part of the Poyang
Lake Plain decreased by 13. 5%. With corrected forecast errors oscillating around the zero axis and a noticeable
improvement in systematic positive errors， the model significantly overestimates the temperature forecast for
high mountain areas. The temperature forecast errors oscillate with the smallest amplitude from August to Octo‐
ber and the largest amplitude in spring and winter. Taking the warming process （May 1-6， 2022） and the strong
cooling process （November 28-December 3， 2022） as examples， the corrected MAE decreased by 18. 2% and
16. 0%， respectively， indicating that the method has achieved stable correction effects during transitional weath‐
er. This composite method has good stability， strong forecast correction ability， easy to promote.
Key words： CMA-GD； hourly； temperature forecast； regression； one-dimensional Kalman； moving average；
error correction