Mountainous winds exhibit strong intermittent， fluctuating， and non-stationary characteristics due to the influence of terrain， resulting in poor observation quality， which makes conventional quality control methods unable to effectively improve their observation quality.To address this issue， a quality control method （VCG） based on variational mode decomposition， convolutional neural networks， and deep learning of gated cyclic units is constructed， and a particle swarm optimization strategy and wind power reconstruction model are introduced to comprehensively improve the quality of observation data.To verify the effectiveness of this method， 10 minute wind speed and direction data of target wind turbines in six complex mountainous wind farms in Jiangxi Ganzhou， Sichuan Guangyuan， Anhui Wuhu， Hubei Huangshi， Henan Pingdingshan， and Guangxi Hezhou in 2016 was quality controlled by VCG and compared with single machine learning method， spatial regression method （SRT）， and inverse distance weighting method （IDW）.The results indicate that VCG method is suitable for quality control of observed wind data in mountainous wind farms， and has a higher error detection rate for suspicious data compared to conventional methods； The controlled data can better restore the observed background field and have a lower error rate when applied to the power generation evaluation business of wind farms； And it has the characteristics of strong terrain adaptability.