Abstract:In order to explore the influence of different time scales data source on groundwater depth prediction and increase the accuracy of depth of groundwater prediction in the freezing and thawing irrigation area, the multivariate time series (CAR) model with monthly, quarterly and annual data were studied, and the differences, including different time scale data source and different input variables were analyzed to decrease the effects of groundwater hysteresis and nonlinear in Yongji irrigation field, Hetao Irrigation Area. The results showed that the CAR model with quarterly data source was obviously better than that with monthly and annual CAR model. The determination coefficient (R2), the Nash-Sutcliffe coefficient (Ens) and the rootmeansquare error (RMSE) of the CAR model with quarterly scale data were 0.936, 0.934 and 0.046m, respectively. Compared with the CAR model with monthly scale data, the R2 and Ens were increased by 1130% and 11.86% and RMSE was decreased by 32.35%. Compared with the CAR model which only considered the freezing and thawing temperature, R2 and Ens of the CAR model considering the whole year temperature and CAR model without temperature were decreased by 0.53%, 0.64% and 2.98%, 3.09%, RMSE was increased by 4.55% and 11.36%. The CAR model with quarterly scale data and only the temperature in freezing and thawing period source was the optimal groundwater predictive model in the region, and R2 was 0.941, Ens was 0.940, and RMSE was 0.044m, with high simulation accuracy, which can provide reference for groundwater depth prediction in freezing-thawing irrigation area.