A systematic procedure is proposed in the present paper to incorporate the options of merging and/or splitting streams from multiple origins in process synthesis. In particular, design problems associated with the heat recovery systems in chemical processes are discussed in detail. First, an input-output system structure is described to facilitate the derivation of the improved LP and MILP models for calculating the design targets corresponding to the operating and capital costs. Next, the construction procedure of a general stream structure is proposed to formulate a NLP model for generating the optimal network configurations. The feasibility and effectiveness of this approach is demonstrated with five examples.