Abstract: In order to accurately and efficiently solve the dynamic reactive power optimization which is taken as mix-integer non-linear programming with strong space-time coupling, a decoupling strategy of relaxation-cluster- ing-correcting is proposed. Firstly, the 24-hour optimal reactive power compensation value of capacitor bank is obtained by relaxing discrete variables and switching operation constraints of the capacitor bank. Secondly, time periods are divided and the actual compensation capacity of capacitor banks is determined based on K-Means clustering. Finally, the dynamic reactive power optimization results are determined by correcting continuous variables. The strategy only needs to solve the nonlinear programming model, which can reduce the solution scale and obtain scheduling results with high satisfaction. In the process of optimization, a self-adaptive learning of multi-mechanism based particle swarm optimization is designed to solve the mode, which has three particle evolution mechanisms with different advantages according to the characteristics of the model. In the iterative process, the execution probability of the three mechanisms is dynamically adjusted to give full play to the advantages of each mechanism, so as to overcome the shortcomings of the traditional particle swarm optimization which is slow in convergence and easy to falling into the local optimal solution. The IEEE 33-bus system is presented to verify effectiveness of the proposed decoupling strategy and algorithm.