Journal of Networking Technology Xinjie Zuoa, Chang Xub, Zhengbo Wangc, Yanfei Long 16 2 2025 https://doi.org/10.6025/jnt/2025/16/2/62-79 https://www.dline.info/jnt/fulltext/v16n2/jntv16n2_3.pdf Under the influence of geographical environment and seasonal differences, wind speed changes show strong characteristics of volatility, intermittency and high variability. To make better use of wind energy in the Kumtag Desert of Xinjiang, this paper proposes a Conv-Informer model and a loss function including a trend penalty factor, combined with a data set partitioning strategy. It researches ultra-short time surface wind prediction in the Kumtag Desert based on deep learning. The results show that the data set partitioning strategy has a crucial impact on the effect of model training. In the results of Conv-Informer model training, the prediction accuracy of the wind speed above 10 m/s within 30 minutes is more than 93%, and the linear trend of wind speed change can be predicted more accurately.