In order to realize the rational distribution for the reactive power between the series and parallel units of unified power quality conditioner a (UPQC) and improve its compensation efficiency, a new method that is capacity optimization and control is promoted. The steady characteristics of UPQC are analyzed by using vector diagram, and the relationship model among capacity, voltage sag/swell amplitude, load capacity, load power factor angel and variation of load voltage phase angle is established. According to the load voltage phase which is controlled after compensation, the online estimation method based on neural network is proposed, which can reduce the real-time compensation power. At last, the circuit simulation result shows that the proposed method can effectively reduce the UPQC design capacity and improve the efficiency of UPQC dynamic compensation. This work is supported by the Scientific Research Fund of State Ethnic Affairs Commission (No. 14BFZ010).