Abstract: The frequent voltage exceeding of the distribution network caused by the high proportion of distributed photovoltaic grid connection is rooted in the time scale mismatch between the rapid fluctuation of photovoltaic output and traditional slow regulation methods. This article proposes a dual layer voltage collaborative control architecture: the upper layer adopts a partition model predictive control, and based on a linearized voltage sensitivity model, the total reactive/active power adjustment of each region is optimized through rolling optimization; The lower layer adopts a consistency algorithm to distribute power instructions among inverters in the region according to the equal incremental rate criterion. The results indicate that this method constrains the voltage within 0.95, 1.05 p.u. throughout the entire time period; In typical over limit events, the total active power reduction is reduced by 66.7% and 25% respectively compared to droop control and centralized MPC, and the regulation cost is reduced by 7.9%, verifying its comprehensive advantages in control accuracy and economy.