Land use is highly dependent on scale and heterogeneous, serving as an important tool to examine land use change by comparing spatial multi-scale cultivated land models through an appropriate simulation model. In the present study, such indicators as the goodness of fit, the spatial auto-correlation of residuals, the number of factors and spatial scale are adopted to compare the variability and accuracy of four spatial multi-scale models: OLS (Ordinary Least Squares), SL (Spatial Lag), SE (Spatial Error) and GWR (Geographically Weighted Regression). According to the global characteristics, there is a positive autocorrelation between the distribution of cropland and the driving factors at multiple spatial scales and they are relatively sensitive to the sampling scale. According to the local characteristics, the smaller the spatial scale, the more accurate the spatial location of regional cropland aggregation. The order of merit of models characterizing the local spatial pattern and distribution pattern of cultivated land at the scale of plateau urban clusters is as follows: GWR, spatial error model, spatial lag model, and OLS. The GWR performs better in characterizing local spatial features, simulation accuracy and driving factor coefficients. Besides, the CLUE-S simulation model improved by GWR Logistic is more accurate in characterizing the spatial pattern of local land use and its distribution pattern at the scale of urban clusters in the plateau region.