Growth and diffusion models have been introduced in such diverse fields such as psychology, economics, marketing, epidemiology, and biology. In this paper, a new nonlinear combinatorial diffusion model is also proposed to remove deficiencies associated with the current diffusion model such as trend curve models, hybrid models. Furthermore the corresponding back propagation learning algorithm is put forward. Theoretical analysis and forecasting examples all show that the new technique has reinforcement learning properties and universalized capabilities. With respect to combined modeling and forecasting of innovation diffusion time series in nonlinear systems, which have some uncertainties, the methods are feasible and effective.[WT5HZ]