In order to improve the non-dominant solution set with better convergence and distributivity of dynamic multi-objective problems, a multi-population decomposition prediction algorithm is proposed. Firstly, an evolutionary vector adaptive generation strategy is proposed, which generates a set of uniform evolutionary vectors based on the solutions of preference sub-problems to ensure the convergence and distribution of the Pareto set. Then, the new non-dominant solution is obtained based on the location of the solution of the subproblem in the target space. Thirdly, a predictive model is adopted to initialize the subpopulation to improve the optimum speed and performance of the algorithm. The experimental results show that compared with five existing algorithms, the proposed algorithm has obvious advantages in convergence and distribution over ten standard dynamic multi-objective problems..