Abstract: Error correction for temperature forecasting is significant for social development and personal life. The U-Net-based models have limitations in correcting meteorological data with geographical differences due to their inherent inductive bias. To address this, a Spatial Attention-based Grid Temperature Correction Network (SAGTCN) was proposed, which uses fully connected networks for grid-point-independent correction and adds spatial attention layers for dynamic regional weighting. On Inner Mongolia's meteorological data, SAGTCN outperforms U-Net and Attention U-Net-based models, with a 5.8% lower Mean Absolute Error (MAE), 14.96% and 10.98% lower Mean Square Error (MSE). In the test of forecasts with 13 - 48 hours lead time, SAGTCN is better in 31 out of 36 lead times, with only 3 close ones, and 2 slightly worse ones.

Key words: air temperature forecasting, error correction, deep learning, spatial attention

摘要： 气温预报误差的订正方法对社会发展和个人生活都具有重要意义。以U-Net为基础的模型受本身归纳偏置限制，对具备显著地理差异的气象数据订正存在一定局限性。为解决该问题，采用全连接网络对格点独立订正，并添加空间注意力层，对不同区域进行权值动态调整，提出了基于空间注意力的格点气温订正网络SAGTCN。在内蒙古自治区气象数据集上，将SAGTCN模型与基于U-Net和Attention U-Net的订正模型进行了对比。实验结果表明，SAGTCN模型的平均绝对误差MAE比二者低5.8%，均方误差MSE分别低14.96%和10.98%。在预报时效为13小时~48小时的对比中，SAGTCN模型的订正效果除在2个时效上略差、3个时效接近外，在其他31个预报时效上都要优于对比模型。

关键词: 气温预报, 误差校正, 深度学习, 空间注意力