To solve the contradiction between population diversity and convergence in particle swarm optimization, an improved particle swarm optimization algorithm, called dynamic multi-population particle swarm optimization with recombined learning and hybrid mutation, is proposed. In the proposed algorithm, a population is divided dynamically and a new particle is reconstructed as a guiding factor. It retains the spatial information of the excellent particles while increasing population diversity. During the execution of the algorithm, a hybrid mutation strategy is applied to adjust the optimal solution. The opposition-based learning and neighborhood-disturbance operations are implemented based on a time-varying probability, which help the particles jump out of the local dilemma quickly, and strengthen good searching in the nearby areas. The effectiveness and superposition effects of several proposed improvement operations compared with several improved particle swarm algorithms based on a set of 14 multi-type benchmark functions are verified. In order to further explore the sensitivity of probability-based hybrid mutation strategy, a large number of simulation experiments are carried out to analyze the mutation mode and parameter settings. The results show that the disturbed extreme strategy has significant advantages. Controlling the learning intensity reasonably can make the opposition-based learning show better performances, furthermore, a suggested value range is given. Finally, experimental results show that the proposed algorithm can get a better balance between the exploitation and exploration for the swarm searching and improve the solution accuracy and convergence performance.