日最高、最低气温的预报是天气预报业务的重要组成部分，其精度提升对保障社会经济活动具有重要意义。针对数值模式在复杂地形区域系统性偏差显著的问题，本研究以湖南省为试验区（具有“凹”字形三级阶梯地貌，涵盖山地、丘陵、平原等多种下垫面类型），提出一种融合地理特征聚类的深度学习改进方案。基于欧洲中期天气预报中心（European Centre for Medium Range Weather Forecasts，ECMWF）模式预报场、中国气象局陆面数据同化系统（CMA Land Data Assimilation System，CLDAS）再分析数据和地理变量，首先构建基准卷积神经网络CNN模型并针对地形特征处理策略设置三类对照实验：方案1（K-means聚类地理变量）、方案2（标准化非聚类地理变量）及方案3（无地理变量）。对比实验表明，方案1 对 24 h 最高/低温度平均绝对误差 MAE 较方案 3 分别降低 4. 7%/9. 4%，较方案 2 预报技巧提升 2. 5%/1. 4%，证实地理特征聚类处理对模型性能具有显著增益。因此，优选方案 1发展了未来 72 h气温预报CNN-TC（Terrain Correction）模型。该模型表现出显著预报优势：相较ECMWF产品，最高/低温MAE降幅 达 23. 5%~37. 3%/20. 8%~26. 9%；较 中 央 气 象 台 指 导 预 报 产 品 SCMOC，最 高/低 温 度 误 差 降 低18. 7%~27. 6%/26. 8%~32. 3%，其中 24 h 预报时效下，最高/低温度空间分布 MAE 区间由 1. 2~5. 8 ℃/0. 8~5. 9 ℃（ECMWF）降低至0. 9~1. 7 ℃/0. 8~1. 7 ℃，区域稳定性大幅提升。分月检验表明，CNN-TC模型在所有月份均保持最优性能，MAE相对降幅覆盖5. 6%~59. 1%（最高温度）和6. 3%~47. 8%（最低温度）。典型强天气过程检验中，模型成功捕捉2022年11月寒潮过程的降温特征，较ECWMF和SCMOC表现均为最优，显示出优异的极端天气应对能力。本研究证实通过深度学习耦合地形特征聚类，可有效解决数值模式在复杂下垫面区域的系统性偏差问题，为山地气候区精细化气象服务提供了可靠的技术方案。

Accurate forecasting of daily maximum（Tmax）and minimum（Tmin）temperatures is essential for meteorological operational，as enhanced precision safeguards socioeconomic stability in agriculture，transportation and public health. To address pronounced systematic biases in numerical weather prediction models over complex terrain-particularly regions with heterogeneous underlying surfaces-this study develops an advanced deep learning framework integrating geographic feature clustering. The study focuses on China’s Hunan Province，characterized by a distinctive concave-shaped，three-tiered topography encompassing mountains，hills，basins，and plains. A baseline convolutional neural network（CNN）was constructed using ECMWF forecast fields，high-resolution CLDAS（China Land Data Assimilation System）reanalysis data，and multi-dimensional geographic variables（elevation，slope，aspect，terrain roughness index）. Three comparative experiments rigorously evaluated terrain-processing efficacy：Method 1 （K-means clustered geographic variables delineating topographic regimes），Method 2（Conventionally standardized non-clustered geographic variables），and Method 3（Geographic variable exclusion as terrain-agnostic control）. Validation confirmed Method 1's superiority，achieving 24- hour mean absolute error（MAE）reductions of 4. 7% for Tmax and 9. 4% for Tmin relative to Method 3，while im‐ proving forecast skill by 2. 5%（Tmax）and 1. 4%（Tmin）compared to Method 2. These statistically significant gains validate the benefits of explicit terrain feature clustering. Building on this approach，the CNN-Terrain Cor‐ rection（CNN-TC）model was developed for 72-hour Tmax/Tmin predictions. CNN-TC framework delivers transformative improvements. Versus ECMWF outputs，Tmax MAE decreased by 23. 5%~37. 3% and Tmin MAE by 20. 8%~26. 9% across 24~72 hour lead times. Compared to operational SCMOC products，errors reduced 18. 7%~27. 6%（Tmax）and 26. 8%~32. 3%（Tmin）. Critically，the model compresses spatial error dispersion，narrowing 24-hour MAE ranges from 1. 2~5. 8 ℃/0. 8~5. 9 ℃（ECMWF）to a stable 0. 9~1. 7 ℃/0. 8~1. 7 ℃ for T max/Tmin-demonstrating breakthrough operational stability. Monthly verification confirmed persistent superiority， with T max MAE reductions of 5. 6%~59. 1% and Tmin improvements of 6. 3% to 47. 8% relative to ECMWF. During the November 2022 cold-air outbreak，the model captured intricate spatiotemporal cooling patterns，outperforming ECMWF and SCMOC with over 30% error reduction，underscoring its capability in extreme weather. This study verifies that deep learning combined with terrain clustering effectively mitigates systematic biases over complex terrain. The CNN-TC framework establishes a robust solution for refined meteorological services in mountainous regions. Cross-regional implementation requires localized hyperparameter optimization and cluster retraining to address spatial heterogeneity in climate regimes and surface properties.