Dynamic multi-objective evolutionary algorithms (DMOEAs) can effectively solve dynamic multi-objective optimization problems (DMOPs) by extracting knowledge from historical environments to predict new solutions. Currently, most existing prediction-based DMOEAs rely heavily on information from the previous two historical environments and lack dynamic maintenance and selective reuse of the extracted historical knowledge. To address this issue, this paper proposes a DMOEA based on adaptive fusion prediction of multi-source knowledge, named MSFP-DMOEA. The framework captures nonlinear change trends in the decision space through a dynamic feature projection network (DFPN) and systematically stores and retrieves optimization patterns from historical environments using a multi-modal knowledge base (MultiModalKB). Finally, a multi-source knowledge fusion prediction mechanism is employed to weightedly fuse the recent environment trend prediction and the historical similar pattern prediction, generating a high-quality initial population. Meanwhile, dynamic maintenance of historical knowledge is achieved through modal update, merging, and elimination mechanisms. Experimental results on multiple benchmark problems and a practical application scenario demonstrate that the MSFP-DMOEA is significantly superior to several current mainstream dynamic multi-objective optimization algorithms in terms of convergence, diversity, and stability.