Abstract: The stress-strain constitutive relation of FRP-confined concrete can be divided into two categories of constraint models under axial load:work hardening and softening. The intrinsic axial compressive stress and strain in the ultimate state constitute the parameter basis of their constitutions. Calculating these constitutive parameters accurately can provide a discrimination criteria for evaluating the performance of FRP-confined concrete structures. The performance of the existing empirical models about compressive strength and ultimate compressive strain of rectangular concrete confined with FRP was evaluated comprehensively, and the results showed that the existing empirical models generally showed poor universality, low prediction accuracy and high dispersion. Aiming at the limitation of traditional prediction models, the prediction models of compressive strength and ultimate compressive strain for both hardening and softening types of rectangular concrete confined with FRP were constructed respectively based on back-propagation artificial neural network. The research results indicated that the neural network models could not only reflect the influence of various control parameters on the axial compressive stress and strain, but also the calculated values based on the neural network models were in better agreement with the experimental values when compared with the existing empirical models; moreover, the deviation and randomness were significantly decreased,so as to ensure the accuracy and stability of prediction results.