A series of creep tests of BSTMUF601 superalloy were carried out at different loads and temperatures to investigate creep behaviors at actual service environment. The diameter correction method was proposed to evaluate true stress and strain approximately for addressing the issue that the decrease of sectional area of specimens. And the θ projection creep constitutive model was used for characterizing creep deformation behaviors considering the advantage of reflecting the deformation process under constant true stress conditions. However, the parameters of creep constitutive model cannot be identified accurately by nonlinear multivariate fitting method under constant load conditions. In this paper, these constitutive parameters were calibrated by BP neural network importing temperature, time, stress and strain evaluated from the above correction method as inputs with back-propagation learning algorithm. Consequently, the calibrated constitutive model is determined, the predicted values coincide well with experimental results and the maximum relative error is less than 12%. Moreover, both the apparent creep stress exponent estimated by θ model, experimental results and the TEM patterns indicated the creep deformation mechanism may be dislocation climb, further indicating the BP neural network method is feasible for predicting complex models.