Abstract: Temporal evolution characteristics of the GRAPES global 4DVar analysis increment in the spectral space is investigated, and the impact on the analysis and forecast is also analyzed when the background error horizontal correlation characteristics at the beginning and end of the data assimilation time window are obviously different. Three types of horizontal correlation model, i.e., the second-order autoregressive model (SOAR), the statistical model from the samples generated by ensemble of data assimilation (EDA), and the blending results of SOAR and EDA, are compared. The results show that the information of synoptic scale in the analysis increment is obviously underestimated by the SOAR model. For the horizontal correlation power spectrum calculated by blending of the SOAR and the EDA, the results show a multi-scale horizontal correlation characteristics, which can better absorb observation information and significantly improve the forecast of geopotential height and temperature in the northern hemisphere. The wind forecast is also improved in the southern hemisphere, while a neutral impact is found in the tropical region. The above results indicate that the merging horizontal correlation scheme developed in this paper is reasonable and practical.