Due to over-exploitation of groundwater in many cities of North China Plain, there is a tendency of lasting decrease in groundwater level, which results in serious problems, such as groundwater exhaustion, land subsidence and seawater intrusion. In order to accurately predict changes of urban groundwater level, based on artificial neural network (ANN) and analysis of multi-scale of wavelet transform (WT), we established a wavelet-ANN conjugate model and test its accuracy to predict groundwater level. Measured data of groundwater level at Pinggu district of Beijing were taken as research objects. We predicted groundwater levels at the district by back propagation (BP) model and hybrid model. Then, we calculated the prediction accuracy by using statistical parameters including root mean square error (RMSE), mean absolute error (MAE) and correlation coefficient (R). Results showed that the MAE of the hybrid model from the first month to the third month was 0.535, 0.598 and 0.634 m, respectively, whereas 0.566, 0.824 and 0.940 m for BP model. The MAE of hybrid model from the first month to the third month was 95%, 73% and 67% of that of BP model, respectively. Comparison of results reveals that the hybrid model has advantages of better prediction accuracy and longer effective prediction duration.