The affecting factors of ultra-tight sandstone reservoir permeability are complex in the Third Member of Xujiahe Formation in Dayi structure of the southern Western Sichuan Depression. Through the analysis and calculation，it is found that the prediction accuracy is not ideal by the various classical permeability prediction models based on the capillary pressure curves，which has certain limitations. This paper analyzed the reasons for the prediction error of six classical models，selected characteristic parameters，and comprehensively considered multiple affecting factors of permeability such as pore throat size and pore throat distribution to improve prediction accuracy of permeability in ultra-tight sandstone reservoirs.On this basis，the method of leave-one-out cross-validation（LooCV）was used to determine the optimal number of latent variables in the models，and the method of partial least squares regression（PLSR）was employed to construct three prediction models for ultra-tight sandstone reservoir permeability，i. e.，Winland-r5（PLSR），Pittman（PLSR），and Swanson （PLSR）. In this way，the problems of the permeability prediction models based on the ordinary least squares（OLS）method can be effectively solved，such as the multicollinearity of many characteristic parameters and the inability of small samples to generalize models. The results reveal that the three permeability prediction models based on PLSR have strong generalization ability，high prediction accuracy，and good applicability in the ultra-tight sandstone reservoirs of the study area.