In cloud manufacturing (CMfg) model, both manufacturing tasks and manufacturing services are in a dynamic environment, therefore the dynamic adaptability of the manufacturing service composition needs to be solved urgently. To address this problem, a theoretical model for manufacturing task and manufacturing service dynamic matching network (DMN) is constructed based on the matching relationships between manufacturing tasks and manufacturing services. Based on this model, a three-phase manufacturing service composition self-adaptive approach (TPMSCSAA) is proposed in this paper. Firstly, by using the load and dynamic QoS evaluated by the load queue model as the optimization goals, the optimal manufacturing service composition is transformed into the shortest path search in the manufacturing service network, and thus dynamic scheduling of manufacturing service composition is realized. Secondly, the changes of manufacturing tasks and manufacturing services are obtained to refresh the manufacturing task network and manufacturing service network in real time. Thirdly, the dynamic scheduling algorithm is triggered to complete the reconstruction of the dynamic matching edges. Finally, the experimental simulation of elevator design service composition is carried out to validate and verify the proposed approach.