Page 191 - 《高原气象》2025年第6期

6 期 周秋雪等：一种适用于复杂地形下最高气温订正的机器学习方法 1599

A Machine Learning Method for Maximum Temperature

Bias Correction in Complex Terrain

1， 2
1， 2
ZHOU Qiuxue ， FENG Liangmin ， CHEN Chaoping ， HU Di 1， 2
1， 2
（1. Sichuan Meteorological Observatory， Chengdu， 610072， Sichuan， China；
2. Heavy Rain and Drought-Flood Disasters in Plateau and Basin Key Laboratory of
Sichuan Province， Chengdu， 610072， Sichuan， China）

Abstract： To enhance the prediction accuracy of the 2-meter maximum temperature in complex terrain areas，
this study developed a gradient modeling approach based on the LightGBM （Light Gradient Boosting Machine，
LGB） algorithm， applied to the Sichuan Basin and its surrounding regions. By selecting and and analyzing multi‐
ple meteorological and topographic factors， an optimized model was constructed. The results demonstrate that：
（1） From January to June 2024， the LightGBM model reduced the mean absolute error by 2. 48 ℃ and improved
the forecast accuracy by 36. 97% compared to EC model. Among them， the improvement effect of the west Sich‐
uan Plateau and Panxi area was the most significant， the accuracy rate increased by 67. 2% and 57. 5%， respec‐
tively.（2） Compared with the existing objective forecast products SPCO and SCMOC， the LightGBM model im‐
proved prediction accuracy by 5. 1% and 10. 3%， respectively. Particularly in the Panxi area and the Sichuan Ba‐
sin， the accuracy at individual stations improved by up to 17. 6% and 23. 4%， respectively.（3）The LightGBM
model reduced the mean absolute error by 2. 05~2. 78 ℃， and increased the accuracy by 31. 1%~41. 0%， with
the most notable enhancement occurring in April.（4）The LightGBM model exhibits strong scalability. Future
work could further improve temperature prediction across Sichuan Province and other regions by incorporating
time-lag effects， spatial neighborhood characteristics， and combining zoning modeling and multi-model integra‐
tion.
Key words： maximum temperature； LightGBM； complex terrain； machine learning