Based on the work platform of Matlab-7.0, the paper predicted the depth of groundwater in Minqin Oasis of Gansu Province with the improved three-layers Back-Propagation Neutral Network (BP Neutral Network) model. The inputting 6 factors include monthly irrigation water volume, outflow of Hongyashan reservoir, and monthly precipitation, evaporation, air temperature, and time sequence; the output factor is the groundwater level in Minqin Oasis. The BP Neural Network Model became more sensitive to the temporal evolution when the time sequence factor was input; The Levenberg-Marquardt algorithm could minimize the bias values, and the use of the Bayesian regularization could optimize the combination of squared errors, weights and the sum of the squared threshold. The modeling results were evaluated with the correlation coefficients, relative error, efficiency index, etc. The results showed that the improved model could improve model's simulation precision and stability.