To solve the distribution network minimum loss reconfiguration problem, we propose an improved dynamic multi-swarm particle swarm optimization (DMS-PSO) in this study. In this method, a set-based updating strategy is introduced to solve the combination problem and a neighborhood's learning strategy is combined to improve the algorithm's local search ability. In addition, to address topological infeasible solutions which take up a large proportion in whole solution space, a repair mechanism is presented. The performance of the proposed method is tested on PG&E 69 nodes and 203 nodes distribution systems and compared to other existing methods. The results show that our method has the highest accuracy and lowest computational time.