A novel dynamic surrogate model based optimization (DSMO) for centralized thermoelectric generation (TEG) system affected by heterogeneous temperature difference (HeTD) is designed to achieve maximum power point tracking (MPPT) in this article. Since HeTD usually results in multiple local maximum power points (LMPPs), DSMO needs to rapidly approximate the global maximum power point (GMPP) instead of being trapped at a low quality LMPP. To avoid a blind search, a radial basis function network is adopted to construct the dynamic surrogate model of input/output feature according to the real-time data of centralized TEG system. Furthermore, a greedy search is adopted to accelerate the convergence based on dynamic surrogate model. Four case studies are undertaken to evaluate the practicability and superiority of the proposed method compared with that of a single LMPP based MPPT method and three common meta-heuristic algorithms. In addition, the implementation feasibility of DSMO is demonstrated by the hardware-in-the-loop (HIL) experiment based on dSpace platform.