Page 204 - 《高原气象》2025年第6期

高 原 气 象 44 卷
1612

Deep Learning Grid-point Temperature Forecasting in Changde

City with Multi-scale Feature Integration

LI Lu 1， 2， 3 ， YOU Xiaoxiong ， CHEN Jingjing ， HU Zhenju ， LU Shu ， LI Qiong 1， 5
1， 2
1， 4
1， 4
1， 4
（1. Hunan Key Laboratory of Meteorological Disaster Prevention and Reduction， Changsha 410118， Hunan， China；
2. Changde Meteorological Bureau of Hunan Province， Changde 415000， Hunan， China；
3. Center for Applied Mathematics of Jiangsu Province， Nanjing University of Information Science and
Technology， Nanjing 210044， Jiangsu， China；
4. Hunan Meteorological Observatory， Changsha 410118， Hunan， China；
5. Hunan Weather Modification Center， Changsha 410118， Hunan， China）

Abstract： This study aims to enhance the accuracy of urban grid temperature forecasts by integrating multi-scale
features from both upper-air and surface levels. Utilizing forecast data from the European Centre for Medium-
Range Weather Forecasts （ECMWF） model and hourly temperature grid observation data from the China Meteo‐
rological Administration's High-Resolution Land Data Assimilation System （HRCLDAS） for the Changde region
from April to September 2021 -2024， a high-resolution Multi-Scale U-Net （MU-NET） model was employed.
Three sets of experiments were designed to develop a deep learning model capable of predicting hourly tempera‐
tures in Changde City for the next 24 hours. The experimental results demonstrate that the MU-NET model，
which integrates surface and upper-air features， exhibits the best correction performance within the study area.
The mean absolute error （MAE） and root mean square error （RMSE） of the MU-NET model were reduced by
22% and 25%， respectively， compared to the ECMWF model. Additionally， the MU-NET model achieved the
lowest MAE values in diurnal variations， particularly excelling in the prediction of daily maximum tempera‐
tures， with an average reduction of over 0. 4 ℃. On a spatial scale， the MU-NET model showed a 60%~80% im‐
provement in forecast skill scores over the ECMWF model in high-altitude regions， with the most significant im‐
provements observed in plain and Dongting Lake areas. During two critical weather events in 2024， the MU-
NET model demonstrated stable forecast performance due to its integration of surface and upper-air features， par‐
ticularly in handling complex weather phenomena. The findings of this study provide new insights for improving
the accuracy of temperature forecasts and offer valuable references for practical meteorological forecasting.
Key words： temperature forecast； deep learning； high resolution； error correction； multi-scale features